AIJ - 4

• (2025). Pandora's box problem with time constraints. AIJ, 349, 104426. (www)

  @article{ , abstract = {The Pandora's Box problem models the search for the best alternative when evaluation is costly. In the simplest variant, a decision maker is presented with n boxes, each associated with a cost of inspection and a hidden random reward. The decision maker inspects a subset of these boxes one after the other, in a possibly adaptive order, and gains the difference between the largest revealed reward and the sum of the inspection costs. Although this classic version is well understood (Weitzman 1979), there is a flourishing recent literature on variants of the problem. Here we introduce a general framework—the Pandora's Box Over Time problem—that captures a wide range of variants where time plays a role, e.g., by constraining the schedules of exploration and influencing costs and rewards. In our framework, boxes have time-dependent rewards and costs, whereas inspection may require a box-specific processing time. Moreover, once a box is inspected, its reward may deteriorate over time. Our main result is an efficient constant-factor approximation to the optimal strategy for the Pandora's Box Over Time problem, which is generally NP-hard to compute. We further obtain improved results for the natural special cases where boxes have no processing time, boxes are available only in specific time slots, or when costs and reward distributions are time-independent (but rewards may still deteriorate after inspection).}, archive = {J_AIJ}, author = {Georgios Amanatidis and Ben Berger and Tomer Ezra and Michal Feldman and Federico Fusco and Rebecca Reiffenhäuser and Artem Tsikiridis}, doi = {10.1016/j.artint.2025.104426}, journal = {Artificial Intelligence}, month = {12}, pages = {104426}, shortjournal = {Artif. Intell.}, title = {Pandora's box problem with time constraints}, volume = {349}, year = {2025}, }
• (2025). Defending a city from multi-drone attacks: A sequential stackelberg security games approach. AIJ, 349, 104425. (www)

  @article{ , abstract = {To counter an imminent multi-drone attack on a city, defenders have deployed drones across the city. These drones must intercept/eliminate the threat, thus reducing potential damage from the attack. We model this as a Sequential Stackelberg Security Game, where the defender first commits to a mixed sequential defense strategy, and the attacker then best responds. We develop an efficient algorithm called S2D2, which outputs a defense strategy. We demonstrate the efficacy of S2D2 in extensive experiments on data from 80 real cities, improving the performance of the defender in comparison to greedy heuristics based on prior works. We prove that under some reasonable assumptions about the city structure, S2D2 outputs an approximate Strong Stackelberg Equilibrium (SSE) with a convenient structure.}, archive = {J_AIJ}, author = {Dolev Mutzari and Tonmoay Deb and Cristian Molinaro and Andrea Pugliese and V.S. Subrahmanian and Sarit Kraus}, doi = {10.1016/j.artint.2025.104425}, journal = {Artificial Intelligence}, month = {12}, pages = {104425}, shortjournal = {Artif. Intell.}, title = {Defending a city from multi-drone attacks: A sequential stackelberg security games approach}, volume = {349}, year = {2025}, }
• (2025). Optimal bailouts and strategic debt forgiveness in financial networks. AIJ, 349, 104424. (www)

  @article{ , abstract = {A financial system is represented by a network, where nodes correspond to banks, and directed labeled edges correspond to debt contracts between banks. Once a payment schedule has been defined, the liquidity of the system is defined as the sum of total payments made in the network. Maximizing systemic liquidity is a natural objective of any financial authority, so, we study the setting where the financial authority offers bailout money to some bank(s) or forgives the debts of others in order to help them avoid costs related to default, and, hence, maximize liquidity. We investigate the approximation ratio provided by the greedy bailout policy compared to the optimal one, and we study the computational hardness of finding the optimal debt-removal and budget-constrained optimal bailout policy, respectively. We also study financial systems from a game-theoretic standpoint. We observe that the removal of some incoming debt might be in the best interest of a bank, if that helps one of its borrowers remain solvent and avoid costs related to default. Assuming that a bank's well-being (i.e., utility) is aligned with the incoming payments they receive from the network, we define and analyze a game among banks who want to maximize their utility by strategically giving up some incoming payments. In addition, we extend the previous game by considering bailout payments. After formally defining the above games, we prove results about the existence and quality of pure Nash equilibria, as well as the computational complexity of finding such equilibria.}, archive = {J_AIJ}, author = {Panagiotis Kanellopoulos and Maria Kyropoulou and Hao Zhou}, doi = {10.1016/j.artint.2025.104424}, journal = {Artificial Intelligence}, month = {12}, pages = {104424}, shortjournal = {Artif. Intell.}, title = {Optimal bailouts and strategic debt forgiveness in financial networks}, volume = {349}, year = {2025}, }
• (2025). The topology of surprise. AIJ, 349, 104423. (www)

  @article{ , abstract = {In this paper we present a topological epistemic logic, with modalities for knowledge (modelled as the universal modality), knowability (represented by the topological interior operator), and unknowability of the actual world. The last notion has a non-self-referential reading (modelled by Cantor derivative: the set of limit points of a given set) and a self-referential one (modelled by Cantor's perfect core of a given set: its largest subset without isolated points, where x is isolated iff { x } is open). We completely axiomatize this logic, showing that it is decidable and pspace -complete, and we apply it to the analysis of a famous epistemic puzzle: the Surprise Exam Paradox.}, archive = {J_AIJ}, author = {Alexandru Baltag and Nick Bezhanishvili and David Fernández-Duque}, doi = {10.1016/j.artint.2025.104423}, journal = {Artificial Intelligence}, month = {12}, pages = {104423}, shortjournal = {Artif. Intell.}, title = {The topology of surprise}, volume = {349}, year = {2025}, }

AMC - 9

• (2026). Improved WENO finite difference method: Treating the multiple discontinuities. AMC, 512, 129762. (www)

  @article{ , abstract = {The classical weighted essentially non-oscillatory method (WENO) performs well in solving hyperbolic conservation laws, but may encounter the numerical instability while treating multiple discontinuities due to the use of equal-width substencils. In order to overcome this problem, we propose an improved WENO finite difference method, namely WENO-rp2. The WENO-rp2 uses r + 2 candidate substencils, which can be divided into two groups, the first group S 1 consists of the classical r r-point substencils and the other group S 2 consists of 2 2-point substencils. Then by introducing a TENO-like switching mechanism, S 2 is used for the final WENO-rp2 reconstruction if the classical one can not handle the multiple discontinuities or is too biased, otherwise S 1 is used. By doing so, the WENO-rp2 maintains the optimal ( 2 r − 1 ) th order of accuracy in the smooth region, avoids the non-physical oscillation near multiple discontinuities, and is more central, but does not significantly increase the computational cost and numerical dissipation.}, archive = {J_AMC}, author = {Jian Ming Wu and Cong Huang}, doi = {10.1016/j.amc.2025.129762}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129762}, shortjournal = {Appl. Math. Comput.}, title = {Improved WENO finite difference method: Treating the multiple discontinuities}, volume = {512}, year = {2026}, }
• (2026). A novel method for manipulating polynomial curves by the Gauss–Legendre control polygon with points interpolating property. AMC, 512, 129760. (www)

  @article{ , abstract = {Gauss–Legendre (GL) curves have recently been introduced, offering a significant advancement over traditional Bézier curves. These curves exhibit remarkable properties, particularly in high-degree cases, where the GL control polygon closely approximates the resulting curve. This is in stark contrast to the much larger Bézier control polygon, which poses practical challenges for curve design. While GL curves retain the endpoint interpolation property, their control over intermediate curve points is limited, as the direct influence of the control polygon is confined to the hodograph at the nodes. To address the challenge of refining high-degree polynomial curves to pass through specific points, we propose a novel method that extends the functionality of GL control polygons to achieve point interpolation. This approach allows for simultaneous control of both the hodograph and the curve itself, ensuring that the curve passes through designated points. We describe the construction of suitable polynomials to serve as weights, facilitating the representation of a curve as a combination of control points. By utilizing these polynomials, we introduce a methodology for directly manipulating both the hodograph and the interpolation points at the nodes, thereby enhancing the precision and flexibility of polynomial curve design using GL polygons.}, archive = {J_AMC}, author = {Soo Hyun Kim and Hwan Pyo Moon and Song-Hwa Kwon}, doi = {10.1016/j.amc.2025.129760}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129760}, shortjournal = {Appl. Math. Comput.}, title = {A novel method for manipulating polynomial curves by the Gauss–Legendre control polygon with points interpolating property}, volume = {512}, year = {2026}, }
• (2026). Stability analysis of extended canonical euler splitting method for nonlinear composite stiff neutral functional differential-algebraic equations. AMC, 512, 129758. (www)

  @article{ , abstract = {In this paper, an extended canonical Euler splitting method is proposed to solve numerically nonlinear composite stiff neutral functional differential-algebraic equations. Under appropriate assumptions, the stability of this method is shown theoretically. Some numerical experiments are provided, which confirm theoretical results.}, archive = {J_AMC}, author = {Weijie Wang and Hongliang Liu and Haodong Li and Shoufu Li}, doi = {10.1016/j.amc.2025.129758}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129758}, shortjournal = {Appl. Math. Comput.}, title = {Stability analysis of extended canonical euler splitting method for nonlinear composite stiff neutral functional differential-algebraic equations}, volume = {512}, year = {2026}, }
• (2026). Hemivariational inequalities and controllability results for second-order non-autonomous evolution system with impulsive effects. AMC, 512, 129750. (www)

  @article{ , abstract = {This research investigates the crucial concept of approximate controllability for a class of complex dynamical systems: second-order neutral non-autonomous evolution systems in Hilbert spaces. These systems, characterized by inherent memory effects (due to the neutral term) and non-smooth behavior (modeled by hemivariational inequalities), pose significant analytical challenges. We begin by rigorously establishing the existence of mild solutions for this intricate system. This crucial step relies on a powerful combination of mathematical tools, including cosine functions, a robust fixed-point technique, and the generalized Clarke’s subdifferential, which effectively handles the non-smoothness arising from the hemivariational inequalities. Building upon this foundation, we delve into the core objective: approximate controllability. This fundamental property explores the system’s ability to be arbitrarily close to any desired target state through judicious selection of control inputs. We derive sufficient conditions for approximate controllability, providing valuable insights into the system’s controllability characteristics. Finally, to underscore the practical significance of our theoretical findings, we present a concrete application demonstrating the developed theory’s efficacy in addressing real-world problems.}, archive = {J_AMC}, author = {Doha A. Kattan and Hasanen A. Hammad}, doi = {10.1016/j.amc.2025.129750}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129750}, shortjournal = {Appl. Math. Comput.}, title = {Hemivariational inequalities and controllability results for second-order non-autonomous evolution system with impulsive effects}, volume = {512}, year = {2026}, }
• (2026). Belonging and refueling: Group dynamics in foraging flocks. AMC, 512, 129748. (www)

  @article{ , abstract = {We extend the Couzin collective motion model with energy–dependent foraging, where agents balance two competing motivations: remaining with the group versus breaking away to refuel. While both behaviors are individually driven, they reflect different priorities–social cohesion versus physiological need. Our results bridge foraging theory with collective motion, revealing how state-dependent behavioral switching produces phase transitions with implications for collective decision-making in natural systems and autonomous swarm design. Local repulsion–orientation–attraction interactions produce group behavior, whereas refueling requires an intentional defection from this emergent cohesion. Our model explores how agents switch between these internally motivated but functionally distinct behaviors and how this switching gives rise to collective phase transitions. Environmental constraints like ground repulsion radius restrict parameter regions supporting coherent motion while preserving energy balance, showing maintained foraging performance despite spatial limitations.}, archive = {J_AMC}, author = {Junhe Qiao and Shannon Dee Algar and Thomas Stemler}, doi = {10.1016/j.amc.2025.129748}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129748}, shortjournal = {Appl. Math. Comput.}, title = {Belonging and refueling: Group dynamics in foraging flocks}, volume = {512}, year = {2026}, }
• (2026). American option pricing with model constrained gaussian process regressions. AMC, 512, 129747. (www)

  @article{ , abstract = {This article introduces a novel method based on Gaussian process regression for pricing American options. The variational partial differential equation (PDE) governing option prices is converted into a non-linear penalized Feynman-Kac equation (PFK). We propose an iterative algorithm to manage the non-linearity of the PFK operator. We sample state variables in the PDE’ s inner domain and on the terminal boundary. At each step, we fit a constrained regression function approximating the option price. This function matches the option payoffs on the boundary sample while satisfying the PFK PDE on the inner sample. The non-linear term in this PDE is frozen and valued with the price estimate from the previous iteration. We adopt a Bayesian framework in which payoffs and the value of the FK PDE in the boundary and inner samples are noised. Assuming the regression function is a Gaussian process, we find a closed-form approximation of option prices. In the numerical illustration, we evaluate American put options in the Heston, Black-Scholes models and in the two-factor Hull-White model.}, archive = {J_AMC}, author = {Donatien Hainaut}, doi = {10.1016/j.amc.2025.129747}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129747}, shortjournal = {Appl. Math. Comput.}, title = {American option pricing with model constrained gaussian process regressions}, volume = {512}, year = {2026}, }
• (2026). On picard-CR iterations involving weak perturbative contraction operators and application to reversible chemical reactions. AMC, 512, 129744. (www)

  @article{ , abstract = {We propose an efficient iterative method called Picard-CR for approximating fixed points under weak perturbative contraction conditions in uniformly convex hyperbolic metric spaces. Theoretical analysis establishes both weak and strong convergence, with performance validated against classical methods (CR, Picard-Noor, and Picard-SP) through numerical experiments. We extend our convergence results to non-expansive and contraction mappings, supported by MATLAB-based visualizations. The iterative scheme is shown to be stable and more efficient, with direct application to computing equilibrium concentrations in reversible chemical reactions. Our findings contribute not only to fixed point theory but also provide practical computational tools for chemical and engineering problems.}, archive = {J_AMC}, author = {Khairul Habib Alam and Aswini Dolai and Yumnam Rohen and Sunil Panday and Shaibal Mani}, doi = {10.1016/j.amc.2025.129744}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129744}, shortjournal = {Appl. Math. Comput.}, title = {On picard-CR iterations involving weak perturbative contraction operators and application to reversible chemical reactions}, volume = {512}, year = {2026}, }
• (2026). Adaptive fuzzy neural network-based smooth-switching gain dynamic surface control for constrained manipulator with error-based nonlinear disturbance observer. AMC, 512, 129743. (www)

  @article{ , abstract = {This paper studies an adaptive fuzzy neural network-based smooth-switching gain dynamic surface control with an error-based nonlinear disturbance observer for the manipulator subject to the full-state constraints and input saturation. First, the smooth-switching gain strategy is proposed to replace the conventional static control gain. This novel strategy can dynamically update the control gains according to the tracking error and its derivative, thereby optimizing tracking performance across different response phases. Second, considering that the conventional nonlinear disturbance observer generates harmful observation peaks when there is a large tracking error, an error-based nonlinear disturbance observer is designed. This observer adaptively adjust the observer gain based on the tracking error, effectively mitigating harmful observation peaks. Furthermore, the adaptive fuzzy neural network strategy is introduced to approximate the modeling uncertainties. Finally, an auxiliary system and an asymmetric time-varying barrier Lyapunov function are established to handle the input saturation and the asymmetric time-varying full-state constraints, respectively. The comparative experiment of a two-degree-of-freedom manipulator validate that the proposed strategy can effectively constrain the state and input of the system within predefined limits. The experimental results further validate that the proposed strategy can effectively optimise the tracking performance and increase the disturbance rejection performance of the system.}, archive = {J_AMC}, author = {Qing Yang and Haisheng Yu and Xiangxiang Meng and Wenqian Yu}, doi = {10.1016/j.amc.2025.129743}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129743}, shortjournal = {Appl. Math. Comput.}, title = {Adaptive fuzzy neural network-based smooth-switching gain dynamic surface control for constrained manipulator with error-based nonlinear disturbance observer}, volume = {512}, year = {2026}, }
• (2026). H∞ guaranteed cost filtering for nonlinear system with limited bandwidth and stochastic hybrid attacks. AMC, 512, 129714. (www)

  @article{ , abstract = {The H ∞ guaranteed cost filter is designed for the system with limited bandwidth in this paper. Due to limited network communication bandwidth, an event-triggered scheme, which fully considers the influence of quantization errors, is proposed in this paper. Under this event-triggered scheme, the burden of network communication can be further reduced. In addition to the event-triggered mechanism, network communication protocol is also an effective method to reduce the burden of network communication. Furthermore, data may be affected by network attacks as it is transmitted through the network channels. Therefore, this paper establishes a Round-Robin communication protocol model under the influence of stochastic hybrid attacks. At last, the effectiveness of the proposed H ∞ guaranteed cost filtering approach is ultimately validated through comprehensive simulation analyses in this paper.}, archive = {J_AMC}, author = {Jian Wang and Zhi-Min Li}, doi = {10.1016/j.amc.2025.129714}, journal = {Applied Mathematics and Computation}, month = {3}, pages = {129714}, shortjournal = {Appl. Math. Comput.}, title = {H∞ guaranteed cost filtering for nonlinear system with limited bandwidth and stochastic hybrid attacks}, volume = {512}, year = {2026}, }

ARTMED - 20

• (2025). Preprocessing narrative texts in electronic medical records to identify hospital adverse events: A scoping review. ARTMED, 170, 103281. (www)

  @article{ , abstract = {Background: Narrative electronic medical records (EMR), which include textual notes created by clinicians within healthcare environments, represent a significant resource for documenting various facets of patient care. This form of text exhibits distinctive characteristics, such as the occurrence of grammatically incorrect sentences, abbreviations, frequent acronyms, specialized characters with particular meanings, negation expressions, and sporadic misspellings. As a result, a primary goal in processing these textual notes is to implement effective preprocessing techniques that enhance data quality and ensure consistency across all entries. Recent advancements in algorithms and methodologies within the fields of natural language processing (NLP), machine learning (ML), and large language models (LLM) have prompted researchers to leverage narrative EMR for the detection of hospital adverse events (HAE). Methods: The scoping review adhered to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines. A scoping review protocol was developed and utilized to guide the research process, clearly outlining the eligibility criteria, information sources, search strategies, data management, selection process, data collection procedures, data items, outcomes and prioritization, data synthesis, and meta-bias considerations. The search strategy was implemented across nine engineering and medical electronic databases. Results: The results have indicated that from a total of 3,264 studies retrieved, 48 unique studies were included in the review. Responses to the research questions were systematically extracted from these studies. The review has identified challenges associated with the preprocessing of narrative texts in EMR for HAE identification. Additionally, three research gaps have been identified: (1) the imperative need for a pipeline to preprocess narrative EMR for the identification of HAE, (2) the necessity for a robust system capable of managing the extensive volume of narrative EMR data, and (3) the requirement for temporal event system, which are essential for effective HAE detection. The study also has underscored the essential role of preprocessing tasks in enhancing the performance of HAE detection. The study has emphasized the importance of extracting N-grams from clinical text, normalizing these N-grams through lemmatization and/or stemming, and establishing semantic feature extraction in preprocessing tasks that significantly affect HAE detection performance. While LLM-based systems naturally incorporate tokenization and normalization processes within their frameworks, it remains crucial to address features that hold semantic relevance to the specific type of HAE during preprocessing. Conclusion: This scoping review has provided valuable insights for researchers focused on HAE detection utilizing narrative EMR data. It has elucidated how preprocessing tasks can elevate the performance of HAE detection and draws attention to neglected research gaps within the field. Addressing these gaps will necessitate further investigation in subsequent research endeavors.}, archive = {J_ARTMED}, author = {Hamed Jafarpour and Guosong Wu and Cheligeer (Ken) Cheligeer and Jun Yan and Yuan Xu and Danielle A. Southern and Cathy A. Eastwood and Yong Zeng and Hude Quan}, doi = {10.1016/j.artmed.2025.103281}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103281}, shortjournal = {Artif. Intell. Med.}, title = {Preprocessing narrative texts in electronic medical records to identify hospital adverse events: A scoping review}, volume = {170}, year = {2025}, }
• (2025). SCLResNet and DSAF: A self-supervised contrastive learning and deep self-attention fusion-based multimodal network for predicting central lymph node metastasis in papillary thyroid carcinoma. ARTMED, 170, 103280. (www)

  @article{ , abstract = {Accurate prediction of central lymph node metastasis (CLNM) in papillary thyroid carcinoma (PTC) is crucial to avoid unnecessary invasive procedures, yet existing models often fall short. We constructed the SCLResNet101 model based on a contrastive learning framework to extract network features of tumor ultrasound (US). SeResnet101 was used to extract network features of peri-vascular adipose tissue (PVAT) from the computed tomography (CT) of C6 (the arterial and venous layers beneath the thyroid). Univariate and multivariate analyses were performed using binary logistic regression to select clinical features. Finally, we constructed a Deep Self-Attention Fusion (DSAF) network to integrate features from these three modalities for CLNM prediction. Univariate and multivariate analyses revealed that Gender, Age, Size of US, and Extrathyroidal Extension (ETE) were independent risk factors for CLNM. In the internal test cohort (I-T), the area under the curve (AUC) of model was 0.863 (95 % CI: 0.779–0.932). In the external test cohort (E-T), the AUC was 0.839 (95 % CI: 0.755–0.905). Compared to all radiologists, the model significantly reduced both false-positive and false-negative rates in both the I-T and E-T. This study incorporates PVAT, which significantly enhances the performance of the multimodal deep learning model and may assist surgeons in making more informed and precise surgical decisions in the treatment of PTC.}, archive = {J_ARTMED}, author = {Shidi Miao and Yuyang Jiang and Wenjuan Huang and Yuxin Jiang and Mengzhuo Sun and Mingxuan Wang and Hongzhuo Qi and Ao Li and Zengyao Liu and Qiujun Wang and Ruitao Wang and Xuemei Ding}, doi = {10.1016/j.artmed.2025.103280}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103280}, shortjournal = {Artif. Intell. Med.}, title = {SCLResNet and DSAF: A self-supervised contrastive learning and deep self-attention fusion-based multimodal network for predicting central lymph node metastasis in papillary thyroid carcinoma}, volume = {170}, year = {2025}, }
• (2025). TSFNet: A Temporal–Spectral fusion network for advanced speech emotion recognition in medical applications. ARTMED, 170, 103279. (www)

  @article{ , abstract = {Speech emotion recognition (SER) is a critical component in enhancing communication systems and human–machine interaction, with significant potential for applications in the medical field. Although existing SER methods that combine temporal and spectral features have achieved notable advancements, they still encounter a big challenge in capturing emotional nuances, which are vital in medical diagnostics and patient care. In this study, we introduce a straightforward yet highly efficient network called TSFNet, which is the Temporal–Spectral Fusion Network via a Large-scale Pre-trained Model. This network is specifically designed to effectively process intricate emotional nuances by seamlessly integrating temporal and spectral information present in speech signals. By leveraging the capabilities of a large-scale pre-trained model, which serves as a powerful plug-and-play component for extracting and learning the temporal characteristics of speech, TSFNet enables a more accurate capture of complex emotional details crucial for medical applications. Extensive experiments are conducted on publicly available datasets, to evaluate the performance of TSFNet. Extensive experiments conducted on six public datasets demonstrate that TSFNet significantly outperforms existing baselines, achieving unweighted accuracies of 95.57% for Savee, 92.67% for Crema-D, 85.71% for IEMOCAP, 100.00% for Tess, 95.86% for Emovo, and 80.43% for Meld. It means that TSFNet has the potential in advancing medical diagnostic tools and patient monitoring systems.}, archive = {J_ARTMED}, author = {Xinran Li and Peilin Huang and Xiaojiang Peng and Feng Sha and Xiaomao Fan and Ye Li}, doi = {10.1016/j.artmed.2025.103279}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103279}, shortjournal = {Artif. Intell. Med.}, title = {TSFNet: A Temporal–Spectral fusion network for advanced speech emotion recognition in medical applications}, volume = {170}, year = {2025}, }
• (2025). Mixture-attention siamese transformer for video polyp segmentation. ARTMED, 170, 103278. (www)

  @article{ , abstract = {Accurate segmentation of polyps from colonoscopy videos is of great significance to polyp treatment and early prevention of colorectal cancer. However, it is challenging due to the difficulties associated with modeling long-range spatio-temporal relationships within a colonoscopy video. In this paper, we address this challenging task with a novel M ixture- A ttention S iamese T ransformer ( MAST ), which explicitly models the long-range spatio-temporal relationships with a mixture-attention mechanism for accurate polyp segmentation. Specifically, we first construct a Siamese transformer architecture to jointly encode paired video frames for their feature representations. We then design a mixture-attention module to exploit the intra-frame and inter-frame correlations, enhancing the features with rich spatio-temporal relationships. Finally, the enhanced features are fed to two parallel decoders for predicting the segmentation maps. Extensive experiments on the large-scale SUN-SEG benchmark demonstrate the superior performance of MAST in comparison with the cutting-edge competitors. Our code is publicly available at https://github.com/Junqing-Yang/MAST .}, archive = {J_ARTMED}, author = {Geng Chen and Junqing Yang and Xiaozhou Pu and Ge-Peng Ji and Huan Xiong and Yongsheng Pan and Hengfei Cui and Yong Xia}, doi = {10.1016/j.artmed.2025.103278}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103278}, shortjournal = {Artif. Intell. Med.}, title = {Mixture-attention siamese transformer for video polyp segmentation}, volume = {170}, year = {2025}, }
• (2025). BPINet: Synchronous blood pressure estimation and user authentication based on ECG and PPG signal with multi-task learning. ARTMED, 170, 103277. (www)

  @article{ , abstract = {As a vital indicator of health, blood pressure is particularly important for elderly individuals with chronic illnesses who live alone. Daily monitoring is essential to prevent hypertension and related complications. Nevertheless, most current home blood pressure monitors depend on cuff-based techniques, which can produce inaccurate results through improper use or cuff placement. Moreover, these devices generally cannot identify the specific user being measured, which hinders the development of personalized long-term health monitoring reports. In this paper, we propose BPINet, a multi-task model based on the Multi-gate Mixture-of-Experts (MMoE) framework that utilizes CNN-BiLSTM to extract features from ECG/PPG signals for simultaneous blood pressure estimation and user authentication (identity recognition). We also compile a dataset of ECG/PPG signals from multiple families, along with their blood pressure measurements, and incorporate it with the University of Queensland Vital Signs Dataset (UQVS) to evaluate the performance of BPINet. On the UQVS dataset, BPINet achieves a 97.54% user identity recognition accuracy. For systolic blood pressure (SBP) estimation, BPINet yields an MAE ± STD of 3.317 ± 5.771 mmHg. For diastolic blood pressure (DBP) estimation, the corresponding values are 2.444 ± 4.147 mmHg. On our customized dataset, BPINet achieves a 94.30% user identity recognition accuracy. For SBP estimation, it yields an MAE ± STD of 2.940 ± 4.753 mmHg. These results meet both the British Hypertension Society (BHS) Grade A standard and the Association for the Advancement of Medical Instrumentation (AAMI) standard. BPINet not only performs blood pressure estimation effectively but also enables simultaneous user identity recognition, facilitating the creation of personalized health records. The experimental results demonstrate the clinical feasibility and effectiveness of our proposed scheme.}, archive = {J_ARTMED}, author = {Xianliang Jiang and Dingxin Yu and Guang Jin and Fei Lei and Weihao Zhang and Xinyan Zhou}, doi = {10.1016/j.artmed.2025.103277}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103277}, shortjournal = {Artif. Intell. Med.}, title = {BPINet: Synchronous blood pressure estimation and user authentication based on ECG and PPG signal with multi-task learning}, volume = {170}, year = {2025}, }
• (2025). The interpretable deep learning framework and validation for seizure detection in pediatric electroencephalography: An improved accuracy and performance analysis. ARTMED, 170, 103276. (www)

  @article{ , abstract = {This study proposes an interpretable deep learning framework and compares the two novel models. A fully convolutional network with squeeze-and-excitation modules (SE-FCN) is designed to enhance spatial sensitivity and retain temporal resolution. In addition, a transformer-based model (TransNet) is developed to capture temporal and channel-wise dependencies via self-attention. These two models output channel saliency weights to the EEG electrode space and generate heatmaps for inferring potential epileptogenic zones. Deep learning primarily adopts convolutional neural networks (CNNs) or sequence generation networks (SGNs) and faces the limitations. For instance, CNN-based models often lack hierarchical modeling and fail to quantify channel-wise contributions, hindering spatial localization. SGN-based models struggle to capture complex spatiotemporal dependencies and typically lack adaptive attention tailored to electroencephalography (EEG) characters. Epileptic seizure detection is vital for effective clinical intervention and existing methods operated as black boxes, limiting clinical interpretability. This study evaluates the models on the CHB-MIT pediatric EEG dataset using a subject-independent cross-validation protocol. SE-FCN achieves an AUC of 0.89 and accuracy of 86.7 %, while TransNet achieves an AUC of 0.92 and accuracy of 86.4 %. Saliency maps from both models demonstrate high consistency and enable categorization of 22 patients into five groups based on inferred seizure origins.}, archive = {J_ARTMED}, author = {Yu Zhou and Yuxin Gao and Qiang Li and Ruiheng Wu and Aiping Yang and Ming-Lang Tseng}, doi = {10.1016/j.artmed.2025.103276}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103276}, shortjournal = {Artif. Intell. Med.}, title = {The interpretable deep learning framework and validation for seizure detection in pediatric electroencephalography: An improved accuracy and performance analysis}, volume = {170}, year = {2025}, }
• (2025). A novel memory interaction neural network for multi-label drug–drug interaction prediction with neighbor importance sampling. ARTMED, 170, 103275. (www)

  @article{ , abstract = {Co-administration of multiple drugs can frequently cause drug–drug interactions (DDIs), including adverse drug reactions (ADRs) that may increase the likelihood of morbidity and mortality. Identifying potential DDIs presents a significant challenge, due to the complexity of pharmacology. Recent advances in knowledge graphs have contributed to DDI prediction by providing a robust framework for representing various relationships between drugs and other entities, such as proteins, diseases, and drug attributes. However, current network-based models often fail to uncover interaction information among DDI triplets, as they typically encode triplets independently. Additionally, uniform sampling methods may overlook differences in neighboring node properties. In this work, we propose a novel memory interaction neural network for DDI prediction, which integrates drug molecular sequences with semantic information from the drug knowledge graph. Specifically, we introduce a neighbor importance sampling strategy that selectively samples highly connected neighbors, improving computational efficiency and reducing noise. We also design a memory interaction module that utilizes multi-head attention mechanisms and deep neural networks to capture interactions among DDI triplets. Experimental evaluation on KEGG and OGB-biokg datasets demonstrates the superiority of our model compared to classical and state-of-the-art methods in predicting DDIs. Datasets and code for this proposed DDIs prediction model are freely accessible at https://github.com/wj1108114106/Multi-label-DDIs .}, archive = {J_ARTMED}, author = {Jing Wang and Runzhi Li and Shuo Zhang and YunLi Xing and Siyu Yan and Lihong Ma}, doi = {10.1016/j.artmed.2025.103275}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103275}, shortjournal = {Artif. Intell. Med.}, title = {A novel memory interaction neural network for multi-label drug–drug interaction prediction with neighbor importance sampling}, volume = {170}, year = {2025}, }
• (2025). Types, functions, and mechanisms of machine learning for personalizing smoking cessation interventions: A systematic scoping review. ARTMED, 170, 103274. (www)

  @article{ , abstract = {Purpose Artificial intelligence can realize personalization. This systematic scoping review provides the types, functions, and mechanisms of machine learning (ML) for personalizing smoking cessation interventions. Methodology We searched fourteen databases including PubMed, CINAHL, EMBASE, the Cochrane Library, IEEE Xplore, PsycINFO, Scopus, Web of Science, AAAI, ACM Digital Library, ArXIV, Mednar, ProQuest, and Science.gov . We selected 98 articles from 4073 records that met the criteria. Two independent reviewers screened and selected the articles. Two reviewers extracted the data using a self-developed data charting form independently. Results The findings are reported in narrative syntheses, tables, and figures. The types of ML included artificial neural networks, Bayesian algorithms, clustering algorithms, decision tree algorithms, deep learning (DL) algorithms, ensemble algorithms, linear classifiers, others, and unspecified. The most common ML technique used was supervised learning (81 %), and the ML functions included (1) message tailoring (17 %), (2) prediction and detection of smoking events (34 %), (3) social media surveillance (14 %), (4) predictive models (24 %), and (5) biomarker analysis (10 %). The ML mechanisms involved the following sequence: data input, data preprocessing, feature extraction and selection, training and validation, and data output. Conclusion This review is the first to describe the potential use of ML for personalizing smoking cessation interventions. We provide recommendations for future research by identifying the limitations and gaps in the studies. Future studies should refine, validate, and test ML models using robust experimental methods to conclude their effectiveness.}, archive = {J_ARTMED}, author = {Yu Jie Xavia Ng and Shing Hui Reina Cheong and Wen Wei Ang and Ying Lau and Siew Tiang Lau}, doi = {10.1016/j.artmed.2025.103274}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103274}, shortjournal = {Artif. Intell. Med.}, title = {Types, functions, and mechanisms of machine learning for personalizing smoking cessation interventions: A systematic scoping review}, volume = {170}, year = {2025}, }
• (2025). Decoding the cortical responses to mechanical wrist perturbations: A two-step shared structure NARX method. ARTMED, 170, 103273. (www)

  @article{ , abstract = {Shared structure nonlinear autoregressive with exogenous input (NARX) model is a promising tool for exploring cortical responses mechanism to external stimuli, essential for advancing our understanding of brain function and developing methods for direct brain information encoding. In this paper, we proposed a two-step method to overcome limitations in existing method, which neglect data relationships and rely on a greedy search for regression terms, leading to less accurate models. In our approach, data from multiple trials are concatenated, and then the orthogonal forward regression (OFR) algorithm identifies model terms in first step, enhancing inter-trial connections and establishing a preliminary model for each subject. Shared model terms across subjects are then used to construct a general target model. Next, non-shared regression terms that best represent population-level information are identified, using adaptive multi-population genetic algorithms, and use to enhance the target models' descriptive power. Simulations results show significant competitiveness in terms of accuracy as compared to other state-of-the-art methods. When applied to real electroencephalography signals under mechanical disturbance, structural and parameter analysis revealed consistent neural response patterns across subjects, with subject-specific responses likely stemming from muscle feedback. Frequency response analysis further suggests that the brain may generate motor inhibition signals based on sensory inputs to maintain a pre-disturbance resting state. These findings provide valuable insights into cortical response mechanisms and have potential implications for future brain information encoding research.}, archive = {J_ARTMED}, author = {Nan Zheng and Yurong Li and Wuxiang Shi and Jiyu Tan}, doi = {10.1016/j.artmed.2025.103273}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103273}, shortjournal = {Artif. Intell. Med.}, title = {Decoding the cortical responses to mechanical wrist perturbations: A two-step shared structure NARX method}, volume = {170}, year = {2025}, }
• (2025). AI-driven dynamic grouping for adaptive clinical trials: Rethinking randomization in precision medicine. ARTMED, 170, 103272. (www)

  @article{ , abstract = {Integrating artificial intelligence into biomedical human subjects research is transforming traditional experimental paradigms. This perspective introduces the concept of “dynamic grouping,” wherein artificial intelligence (AI) systems continuously reassign participants across experimental conditions based on real-time biomarker data and clinical response patterns. Unlike traditional biomedical research designs that rely on fixed treatment and control groups, dynamic grouping allows participant assignments to evolve throughout the study. We examine the ethical implications, methodological challenges, and research opportunities associated with this paradigm, particularly in clinical trials, precision medicine, and digital therapeutics. To support this analysis, we present three computational simulations that quantify its impact: (i) a heterogeneity simulation demonstrating how patient variability affects the advantage of dynamic grouping, (ii) a statistical power analysis showing potential sample size reductions in adaptive designs, and (iii) a clinical outcome distribution analysis highlighting how dynamic grouping reduces negative treatment outcomes and optimizes patient responses. Our findings suggest that dynamic grouping can improve treatment effectiveness, enhance resource allocation, and increase statistical efficiency, although it also raises new challenges for causal inference, informed consent, and regulatory oversight. As AI continues to reshape medical research, adapting ethical and methodological frameworks will be essential for its responsible implementation.}, archive = {J_ARTMED}, author = {Madhur Mangalam}, doi = {10.1016/j.artmed.2025.103272}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103272}, shortjournal = {Artif. Intell. Med.}, title = {AI-driven dynamic grouping for adaptive clinical trials: Rethinking randomization in precision medicine}, volume = {170}, year = {2025}, }
• (2025). H-SynEx: Using synthetic images and ultra-high resolution ex vivo MRI for hypothalamus subregion segmentation. ARTMED, 170, 103271. (www)

  @article{ , abstract = {The hypothalamus is a small structure located in the center of the brain and is involved in significant functions such as sleeping, temperature, and appetite control. Various neurological disorders are also associated with hypothalamic abnormalities. Automated image analysis of this structure from brain MRI is thus highly desirable to study the hypothalamus in vivo . However, most of the automated segmentation tools currently available focus exclusively on T1w images. In this study, we introduce H-SynEx, a machine learning method for automated segmentation of hypothalamic subregions that generalizes across different MRI sequences and resolutions without retraining. H-synEx was trained with synthetic images built from label maps derived from ultra-high resolution ex vivo MRI scans, allowing finer-grained manual segmentation when compared with 1 mm isometric in vivo images. We validated our method using Dice Coefficient (DSC) and Average Hausdorff distance (AVD) across in vivo images from six different datasets with six different MRI sequences (T1, T2, proton density, quantitative T1, fractional anisotropy, and FLAIR). Statistical analysis compared hypothalamic subregion volumes in controls, Alzheimer’s disease (AD), and behavioral variant frontotemporal dementia (bvFTD) subjects using the Area Under the Receiver Operating Characteristic curve (AUROC) and the Wilcoxon rank sum test. Our results show that H-SynEx successfully leverages information from ultra-high resolution scans to segment in vivo from different MRI sequences. Our automated segmentation was able to discriminate controls versus patients with Alzheimer’s disease on FLAIR images with 5 mm spacing. H-SynEx is openly available at https://github.com/liviamarodrigues/hsynex .}, archive = {J_ARTMED}, author = {Livia Rodrigues and Martina Bocchetta and Oula Puonti and Douglas Greve and Ana Carolina Londe and Marcondes França and Simone Appenzeller and Juan Eugenio Iglesias and Leticia Rittner}, doi = {10.1016/j.artmed.2025.103271}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103271}, shortjournal = {Artif. Intell. Med.}, title = {H-SynEx: Using synthetic images and ultra-high resolution ex vivo MRI for hypothalamus subregion segmentation}, volume = {170}, year = {2025}, }
• (2025). Mining multi-electrode and multi-wave electroencephalogram based time-interval temporal patterns for improved classification capabilities and explainability. ARTMED, 170, 103269. (www)

  @article{ , abstract = {Brain-computer interface (BCI) systems, and particularly electroencephalogram (EEG) based BCI systems, have become more widely used in recent years and are utilized in various applications and domains ranging from medicine and marketing to games and entertainment. While different algorithms have been used to analyze EEG data and enable its classification, existing algorithms have two main drawbacks; both their classification and explainability capabilities are limited. Lacking in explainability, they cannot indicate which electrodes and waves led to a classification decision or explain how areas and frequencies of the brain's activity correlate to a specific task. In this study, we propose a novel extension for the time-interval temporal patterns mining algorithms aimed at enhancing the data mining process by enabling a richer set of patterns to be learned from the EEG data, thereby contributing to improved classification and explainability capabilities. The extended algorithm is designed to capture and leverage the unique nature of EEG data by decomposing it into different brain waves and modeling the relations among them and between different electrodes. Our evaluation of the proposed extended algorithm on multiple learning tasks and three EEG datasets demonstrated the extended algorithm's ability to mine richer patterns that improve the classification performance by 4–11 % based on the Area-Under the receiver operating characteristic Curve (AUC) metric, compared to the original version of the algorithm. Moreover, the algorithm was shown to shed light on the areas and frequencies of the brain's activity that are correlated with specific tasks.}, archive = {J_ARTMED}, author = {Ofir Landau and Nir Nissim}, doi = {10.1016/j.artmed.2025.103269}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103269}, shortjournal = {Artif. Intell. Med.}, title = {Mining multi-electrode and multi-wave electroencephalogram based time-interval temporal patterns for improved classification capabilities and explainability}, volume = {170}, year = {2025}, }
• (2025). A survey for large language models in biomedicine. ARTMED, 170, 103268. (www)

  @article{ , abstract = {Recent breakthroughs in large language models (LLMs) offer unprecedented natural language understanding and generation capabilities. However, existing surveys on LLMs in biomedicine often focus on specific applications or model architectures, lacking a comprehensive analysis that integrates the latest advancements across various biomedical domains. This review, based on an analysis of 484 publications sourced from databases including PubMed, Web of Science, and arXiv, provides an in-depth examination of the current landscape, applications, challenges, and prospects of LLMs in biomedicine, distinguishing itself by focusing on the practical implications of these models in real-world biomedical contexts. Firstly, we explore the capabilities of LLMs in zero-shot learning across a broad spectrum of biomedical tasks, including diagnostic assistance, drug discovery, and personalized medicine, among others, with insights drawn from 137 key studies. Then, we discuss adaptation strategies of LLMs, including fine-tuning methods for both uni-modal and multi-modal LLMs to enhance their performance in specialized biomedical contexts where zero-shot fails to achieve, such as medical question answering and efficient processing of biomedical literature. Finally, we discuss the challenges that LLMs face in the biomedicine domain including data privacy concerns, limited model interpretability, issues with dataset quality, and ethics due to the sensitive nature of biomedical data, the need for highly reliable model outputs, and the ethical implications of deploying AI in healthcare. To address these challenges, we also identify future research directions of LLM in biomedicine including federated learning methods to preserve data privacy and integrating explainable AI methodologies to enhance the transparency of LLMs. As this field of LLM rapidly evolves, continued research and development are essential to fully harness the capabilities of LLMs in biomedicine while ensuring their responsible and effective deployment.}, archive = {J_ARTMED}, author = {Chong Wang and Mengyao Li and Junjun He and Zhongruo Wang and Erfan Darzi and Zan Chen and Jin Ye and Tianbin Li and Yanzhou Su and Jing Ke and Kaili Qu and Shuxin Li and Yi Yu and Pietro Liò and Tianyun Wang and Yu Guang Wang and Yiqing Shen}, doi = {10.1016/j.artmed.2025.103268}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103268}, shortjournal = {Artif. Intell. Med.}, title = {A survey for large language models in biomedicine}, volume = {170}, year = {2025}, }
• (2025). High-quality triage and diagnosis of gynecological diseases via artificial intelligence. ARTMED, 170, 103267. (www)

  @article{ , abstract = {Timely detection and diagnosis of diseases are key elements of an efficient healthcare system. In recent years, artificial intelligence (AI) has played an increasingly important role in improving the accuracy and efficiency of disease diagnosis in clinical practice. However, most existing AI systems for disease diagnosis have focused on either classifying patients into broad disease categories or diagnosing a specific disease, leaving a gap in the development of a coherent AI system for both triage and diagnosis in a department of a general hospital. In this study, we fill this gap with SmartGyne, an advanced AI system that can achieve high-quality triage and diagnosis for a full spectrum of gynecological diseases. By extracting useful clinical evidence for diagnosis from a large amount of electronic medical records, SmartGyne establishes an effective framework to integrate real-world clinical evidence and knowledge into a coherent AI system that can effectively handle a full spectrum of complex diseases in a department of a general hospital. Validation experiments demonstrated that SmartGyne achieved an overall accuracy of 80.1 % in triage for gynecological diseases, and 99.4 % in diagnosis for a gynecological subspecialty. In comparison with human physicians, SmartGyne showed competitive triage and diagnostic performance, and improved consultation efficiency and accuracy for physicians with limited specialized experience. These results show that SmartGyne achieves high-quality triage and diagnosis, holding the potential to improve the efficiency of the healthcare system in China, as well as other countries lacking professional gynecologists.}, archive = {J_ARTMED}, author = {Linru Fu and Che Wang and Zhaoyang Liu and Changzai Pan and Zhe Du and Zhijing Sun and Lan Zhu and Ke Deng}, doi = {10.1016/j.artmed.2025.103267}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103267}, shortjournal = {Artif. Intell. Med.}, title = {High-quality triage and diagnosis of gynecological diseases via artificial intelligence}, volume = {170}, year = {2025}, }
• (2025). Improving prototypical parts abstraction for case-based reasoning explanations designed for the kidney stone type recognition. ARTMED, 170, 103266. (www)

  @article{ , abstract = {The in-vivo identification of the kidney stone types during an ureteroscopy would be a major medical advance in urology, as it could reduce the time of the tedious renal calculi extraction process, while diminishing infection risks. Furthermore, such an automated procedure would make possible to prescribe anti-recurrence treatments immediately. Nowadays, only few experienced urologists are able to recognize the kidney stone types in the images of the videos displayed on a screen during the endoscopy. This visual recognition by urologists is also highly operator dependent. Thus, several deep learning (DL) models have recently been proposed to automatically recognize the kidney stone types using ureteroscopic images. However, these DL models are of black box nature and do not establish the relationship of the visual features they used to take the decision with the color, texture and morphological features visually analyzed in biological laboratories to determine the type of extracted kidney stone fragments using the reference morphoconstitutional analysis (MCA) procedure. This contribution proposes a case-based reasoning DLmodel which uses prototypical parts (PPs) and generates local and global descriptors. The PPs encode for each class (i.e., kidney stone type) visual feature information (hue, saturation, intensity and textures) similar to that used by biologists during MCA. The PPs are optimally generated due a new loss function used during the model training. Moreover, the local and global descriptors of PPs allow to explain the decisions (“what” information, “where in the images”) in an understandable way for biologists and urologists. The proposed DL model has been tested on a database including images of the six most widespread kidney stone types in industrialized countries. The overall average classification accuracy was 90 . 37 ± 0 . 6 % . When comparing this results with that of the eight other DL models of the kidney stone state-of-the-art, it can be seen that the valuable gain in explanability was not reached at the expense of accuracy which was even slightly increased with respect to that ( 88 . 2 ± 2 . 1 % ) of the best method of the literature. These promising and interpretable results also encourage urologists to put their trust in AI-based solutions.}, archive = {J_ARTMED}, author = {Daniel Flores-Araiza and Francisco Lopez-Tiro and Clément Larose and Salvador Hinojosa and Andres Mendez-Vazquez and Miguel Gonzalez-Mendoza and Gilberto Ochoa-Ruiz and Christian Daul}, doi = {10.1016/j.artmed.2025.103266}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103266}, shortjournal = {Artif. Intell. Med.}, title = {Improving prototypical parts abstraction for case-based reasoning explanations designed for the kidney stone type recognition}, volume = {170}, year = {2025}, }
• (2025). Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions. ARTMED, 170, 103265. (www)

  @article{ , abstract = {Recent advancements in deep learning have significantly revolutionized the field of clinical diagnosis and treatment, offering novel approaches to improve diagnostic precision and treatment efficacy across diverse clinical domains, thus driving the pursuit of precision medicine. The growing availability of multi-organ and multimodal datasets has accelerated the development of large-scale Medical Multimodal Foundation Models (MMFMs). These models, known for their strong generalization capabilities and rich representational power, are increasingly being adapted to address a wide range of clinical tasks, from early diagnosis to personalized treatment strategies. This review offers a comprehensive analysis of recent developments in MMFMs, focusing on three key aspects: datasets, model architectures, and clinical applications. We also explore the challenges and opportunities in optimizing multimodal representations and discuss how these advancements are shaping the future of healthcare by enabling improved patient outcomes and more efficient clinical workflows.}, archive = {J_ARTMED}, author = {Kai Sun and Siyan Xue and Fuchun Sun and Haoran Sun and Yu Luo and Ling Wang and Siyuan Wang and Na Guo and Lei Liu and Tian Zhao and Xinzhou Wang and Lei Yang and Shuo Jin and Jun Yan and Jiahong Dong}, doi = {10.1016/j.artmed.2025.103265}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103265}, shortjournal = {Artif. Intell. Med.}, title = {Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions}, volume = {170}, year = {2025}, }
• (2025). An end-to-end solution for out-of-hospital emergency medical dispatch triage based on multimodal and continual deep learning. ARTMED, 170, 103264. (www)

  @article{ , abstract = {The objective of this study was to build a multimodal, multitask predictive model—named E2eDeepEMC 2 —to improve out-of-hospital emergency incident severity assessments while coping with shifts in data distributions over time. We drew on 2 054 694 independent incidents recorded by the Valencian emergency medical dispatch service between 2009 and 2019 (excluding 2013), combining demographic, temporal, clinical and free-text inputs. To handle temporal drift, our model integrates continual learning strategies and comprises three encoder modules (for context, clinical data and text), whose outputs are merged to predict the life-threatening level, admissible response delay and emergency system jurisdiction. Compared with the Valencian Region’s existing in-house triage protocol, E2eDeepEMC 2 achieved absolute F1-score gains of 18.46% for life-threatening level, 25.96% for response delay and 3.63% for jurisdiction. Compared to non-continual learning baselines, it also outperformed them by 3.04%, 9.66% and 0.58%, respectively. Deployment of E2eDeepEMC 2 is currently underway in the Valencian Region, underscoring its practical impact on real-world emergency dispatch decision-making.}, archive = {J_ARTMED}, author = {Pablo Ferri and Carlos Sáez and Antonio Félix-De Castro and Purificación Sánchez-Cuesta and Juan M. García-Gómez}, doi = {10.1016/j.artmed.2025.103264}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103264}, shortjournal = {Artif. Intell. Med.}, title = {An end-to-end solution for out-of-hospital emergency medical dispatch triage based on multimodal and continual deep learning}, volume = {170}, year = {2025}, }
• (2025). LoRA-PT: Low-rank adapting UNETR for hippocampus segmentation using principal tensor singular values and vectors. ARTMED, 170, 103254. (www)

  @article{ , abstract = {The hippocampus is an important brain structure involved in various psychiatric disorders, and its automatic and accurate segmentation is vital for studying these diseases. Recently, deep learning-based methods have made significant progress in hippocampus segmentation. However, training deep neural network models requires substantial computational resources, time, and a large amount of labeled training data, which is frequently scarce in medical image segmentation. To address these issues, we propose LoRA-PT, a novel parameter-efficient fine-tuning (PEFT) method that transfers the pre-trained UNETR model from the BraTS2021 dataset to the hippocampus segmentation task. Specifically, LoRA-PT divides the parameter matrix of the transformer structure into three distinct sizes, yielding three third-order tensors. These tensors are decomposed using tensor singular value decomposition to generate low-rank tensors consisting of the principal singular values and vectors, with the remaining singular values and vectors forming the residual tensor. During fine-tuning, only the low-rank tensors (i.e., the principal tensor singular values and vectors) are updated, while the residual tensors remain unchanged. We validated the proposed method on three public hippocampus datasets, and the experimental results show that LoRA-PT outperformed state-of-the-art PEFT methods in segmentation accuracy while significantly reducing the number of parameter updates. Our source code is available at https://github.com/WangangCheng/LoRA-PT/tree/LoRA-PT .}, archive = {J_ARTMED}, author = {Guanghua He and Wangang Cheng and Hancan Zhu and Gaohang Yu}, doi = {10.1016/j.artmed.2025.103254}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103254}, shortjournal = {Artif. Intell. Med.}, title = {LoRA-PT: Low-rank adapting UNETR for hippocampus segmentation using principal tensor singular values and vectors}, volume = {170}, year = {2025}, }
• (2025). MMSupcon: An image fusion-based multi-modal supervised contrastive method for brain tumor diagnosis. ARTMED, 170, 103253. (www)

  @article{ , abstract = {The diagnosis of brain tumors is pivotal for effective treatment, with MRI serving as a commonly used non-invasive diagnostic modality in clinical practices. Fundamentally, brain tumor diagnosis is a type of pattern recognition task that requires the integration of information from multi-modal MRI images. However, existing fusion strategies are hindered by the scarcity of multi-modal imaging samples. In this paper, we propose a new training paradigm tailored for the scenario of multi-modal imaging in brain tumor diagnosis, called multi-modal supervised contrastive learning method (MMSupcon). This method significantly enhances diagnostic accuracy through two key components: multi-modal medical image fusion and multi-modal supervised contrastive loss. First, the fusion component integrates complementary imaging modalities to generate information-rich samples. Second, by introducing fused samples to guide original samples in learning feature consistency or inconsistency among classes, our loss component effectively preserves the integrity of cross-modal information while maintaining the distinctiveness of individual modalities. Finally, MMSupcon is validated on a real-world brain tumor dataset collected from Beijing Tiantan Hospital, achieving state-of-the-art performance. Furthermore, additional experiments on two public BraTS glioma classification datasets also demonstrate our substantial performance improvements. The source code is released at https://github.com/hywang02/MMSupcon .}, archive = {J_ARTMED}, author = {Haoyu Wang and Jing Zhang and Siying Wu and Haoran Wei and Xun Chen and Yunwei Ou and Xiaoyan Sun}, doi = {10.1016/j.artmed.2025.103253}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103253}, shortjournal = {Artif. Intell. Med.}, title = {MMSupcon: An image fusion-based multi-modal supervised contrastive method for brain tumor diagnosis}, volume = {170}, year = {2025}, }
• (2025). Pathway information on methylation analysis using deep neural network (PROMINENT): An interpretable deep learning method with pathway prior for phenotype prediction using gene-level DNA methylation. ARTMED, 170, 103236. (www)

  @article{ , abstract = {Background DNA methylation is a key epigenetic marker that influences gene expression and phenotype regulation, and is affected by both genetic and environmental factors. Traditional linear regression methods such as elastic nets have been employed to assess the cumulative effects of multiple DNA methylation markers on phenotypes. However, these methods often fail to capture the complex nonlinear nature of the data. Recent deep learning approaches, such as MethylNet, have improved the prediction accuracy but lack interpretability and efficiency. Findings To address these limitations, we introduced P athway Info r mati o n on M ethylat i on Analysis using a Deep Ne ural N e t work (PROMINENT), a novel interpretable deep learning method that integrates gene-level DNA methylation data with biological pathway information for phenotype prediction. PROMINENT enhances interpretability and prediction accuracy by incorporating gene- and pathway-level priors from databases such as Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG). It employs SHapley Additive exPlanations (SHAP) to prioritize significant genes and pathways. Evaluated across various datasets, childhood asthma, idiopathic pulmonary fibrosis (IPF), and first-episode psychosis (FEP)—PROMINENT consistently outperformed existing methods in terms of prediction accuracy and computational efficiency. PROMINENT also identified crucial genes and pathways involved in disease mechanisms. Conclusions PROMINENT represents a significant advancement in leveraging DNA methylation data for phenotype prediction, offering both high accuracy and interpretability within reasonable computational time. This method holds promise for elucidating the epigenetic underpinnings of complex diseases and enhancing the utility of DNA methylation data in biomedical research.}, archive = {J_ARTMED}, author = {Soyeon Kim and Laizhi Zhang and Yidi Qin and Rebecca I. Caldino Bohn and Hyun Jung Park}, doi = {10.1016/j.artmed.2025.103236}, journal = {Artificial Intelligence in Medicine}, month = {12}, pages = {103236}, shortjournal = {Artif. Intell. Med.}, title = {Pathway information on methylation analysis using deep neural network (PROMINENT): An interpretable deep learning method with pathway prior for phenotype prediction using gene-level DNA methylation}, volume = {170}, year = {2025}, }

ASOC - 279

• (2025). Deeper-PINNs: Unlocking the power of deep physics-informed neural networks. ASOC, 185, 114048. (www)

  @article{ , abstract = {Physics-Informed Neural Networks (PINNs) have emerged as a promising framework for solving partial differential equations (PDEs) and have garnered significant attention across industrial and scientific domains. However, their effectiveness is often constrained by limited approximation capacity and performance degradation in deep network structures. In this work, we propose the Deeper Physics-Informed Neural Network (Deeper-PINN), a novel architecture designed to address these challenges. The Deeper-PINN incorporates element-wise multiplication operations into the PINN structure, which effectively mitigates the initialization pathologies of PINNs and enables the utilization of deeper network structures. Additionally, this operation projects features into high-dimensional, nonlinear spaces, thereby enhancing the approximation capacity of PINNs. The proposed architecture is evaluated on multiple benchmark problems, demonstrating that Deeper-PINNs can effectively leverage deep neural network structures while maintaining high parameter efficiency. The complete codes of the experiments can be found on https://github.com/flongjiang/Deeper-PINNs}, archive = {J_ASOC}, author = {Feilong Jiang and Xiaonan Hou and Jianqiao Ye and Min Xia}, doi = {10.1016/j.asoc.2025.114048}, journal = {Applied Soft Computing}, month = {12}, pages = {114048}, shortjournal = {Appl. Soft. Comput.}, title = {Deeper-PINNs: Unlocking the power of deep physics-informed neural networks}, volume = {185}, year = {2025}, }
• (2025). Wind turbine blade damage identification using lightweight YOLO11 with multi-scale dilated attention. ASOC, 185, 114042. (www)

  @article{ , abstract = {In recent years, companies such as Siemens and GE have frequently experienced wind turbine shutdowns due to blade damage and fractures, resulting in annual losses of up to $4.6 billion. As a critical component of wind turbines, blades have a high failure rate, which directly affects the safety and reliability of wind turbine operations. However, accurately identifying wind turbine blade damage is a significant challenge, mainly due to the scarcity of damage samples, the small scale of blade damage which is difficult to detect, and the interference from external environmental factors. To address these issues, this paper proposes a high-accuracy and low-consumption damage identification method using SlimNeck-structured YOLO11 with Multi-Scale Dilated Attention (SNMSDA-YOLO11). First, to overcome the issue of limited damage samples from wind turbines, three data augmentation methods are used, including random noise augmentation, image sharpening augmentation, and saturation adjustment augmentation. Second, to solve the problem of adaptive feature extraction for damage at various scales, the multi-scale dilated attention mechanism is integrated into the YOLO11 algorithm. This enhances its ability to extract damage features across different forms of blade damage, thereby improving detection performance. Finally, a slim-neck structure is employed to optimize the model, ensuring accurate damage identification while significantly improving the computational speed and efficiency. Experimental results demonstrate that the proposed SNMSDA-YOLO11 effectively utilizes attention mechanisms to learn multi-scale damage features of wind turbine blades, achieving an accuracy of 94.9 %. Compared to existing methods, this method holds significant potential for large-scale applications in wind turbine blade damage identification.}, archive = {J_ASOC}, author = {Wenjing Xu and Jiachi Yao and Yanxue Wang and Chao Liu and Xinyu Liu and Dongxiang Jiang}, doi = {10.1016/j.asoc.2025.114042}, journal = {Applied Soft Computing}, month = {12}, pages = {114042}, shortjournal = {Appl. Soft. Comput.}, title = {Wind turbine blade damage identification using lightweight YOLO11 with multi-scale dilated attention}, volume = {185}, year = {2025}, }
• (2025). Tea leaf disease detection using UAV remote sensing images based on low-light aware and spatial reconstruction. ASOC, 185, 114034. (www)

  @article{ , abstract = {Tea leaf diseases in unmanned aerial vehicle (UAV) remote sensing images under low-light conditions present various morphologies. Existing methods do not accurately detect these tea leaf diseases. This study proposes a detection method for tea leaf diseases in low-light UAV remote sensing images based on the low-light aware and spatial reconstruction network (LA-SRNet). The low-light aware encoder in LA-SRNet guides the extraction of light invariant features and the noise distribution. The multiscale spatial reconstruction decoder in LA-SRNet achieves image reconstruction. LA-SRNet enhances the feature learning ability of the detection network for disease spots of different scales and morphologies in low-light UAV remote sensing images and alleviating the impact of low light on detection results. The high-frequency feature enhancement block in the tea disease spot detection network (TDSDNet) constructed in this study enhances the high-frequency information of feature maps, improving the detection network’s ability to learn the detailed features of disease spots with different morphologies and reducing the number of missed detections. The multiscale mask attention sub-network in TDSDNet captures the location information of diseased tea leaves, reducing false detections. This study also uses the multiscale mask loss function to suppress the background and highlight the foreground. Experimental results show that the proposed method achieves precision, recall, and mean average precision (mAP)@0.5 of 68.5, 64.8, and 70.1 %, respectively, for the detection of tea leaf diseases in low-light UAV remote sensing images, showing higher values than existing detection methods. The proposed detection network has a mAP value of 1.9 % higher than that of the most advanced detection networks and 3.2 % higher than the baseline network.}, archive = {J_ASOC}, author = {Gensheng Hu and Jiahang Zou}, doi = {10.1016/j.asoc.2025.114034}, journal = {Applied Soft Computing}, month = {12}, pages = {114034}, shortjournal = {Appl. Soft. Comput.}, title = {Tea leaf disease detection using UAV remote sensing images based on low-light aware and spatial reconstruction}, volume = {185}, year = {2025}, }
• (2025). Knowledge-augmented large language model-based framework for adverse drug reaction analysis. ASOC, 185, 114025. (www)

  @article{ , abstract = {Transformer-based pre-trained large language models and their domain-specific variants have achieved impressive performance in natural language processing tasks across various fields, but the static internal prior knowledge limits their capability to adapt to the dynamically evolving knowledge landscape and effectively address knowledge-intensive problems within health informatics, such as identifying adverse drug reactions from online text data generated by the public. Additionally, the non-standardized identification results hinder researchers from conducting in-depth analyzes and drawing meaningful insights. We propose a novel knowledge-augmented large language model-based framework, which incorporates a fine-tuned biomedical large language model for adverse drug reaction identification. External trustworthy knowledge resources are retrieved and integrated into the framework through a unified feature embedding mechanism, systematically enhancing both adverse drug reaction identification and medical terminology standardization. To the best of our knowledge, this study marks initial effort to introduce knowledge augmentation techniques to adverse drug reaction identification and medical terminology standardization. Comprehensive experiments on public benchmark datasets demonstrate the effectiveness of the framework, which outperforms state-of-the-art large language model-based competitors across all evaluation metrics and shows over a 5 % improvement compared to the second-best experimental results.}, archive = {J_ASOC}, author = {Haochen Zou and Yongli Wang and Kun Xiang and Anqi Huang}, doi = {10.1016/j.asoc.2025.114025}, journal = {Applied Soft Computing}, month = {12}, pages = {114025}, shortjournal = {Appl. Soft. Comput.}, title = {Knowledge-augmented large language model-based framework for adverse drug reaction analysis}, volume = {185}, year = {2025}, }
• (2025). A partial load distribution aware cluster migration approach for stable control plane in software-defined wireless sensor networks. ASOC, 185, 114023. (www)

  @article{ , abstract = {The utilization of a distributed architecture in SDN-enabled wireless sensor networks (WSNs) provides significant advantages in terms of reliability and scalability. This architecture uses multiple controllers to provide control services near the data plane to enable prompt responses through control messages. However, the placement of controllers and the balancing of their loads present substantial challenges in SDN-enabled WSNs due to the dynamic nature of devices in the data plane. An articulation point method is used to identify the optimal number of controllers to address the controller placement. Whereas, the nature-inspired seagull optimization approach is considered to place them in the control plane efficiently by considering crucial factors such as the synchronization overhead and network latency. Furthermore, to address load balancing, a novel approach called partial load distribution-based cluster migration (PLD-CM) is proposed. This approach identifies the best-suited assistant controllers of low utilization to reduce the load of overloaded controllers. Subsequently, appropriate control nodes are identified for partial migration in accordance with the load of the assistant controller. To evaluate the performance of the proposed PLD-CM scheme is implemented with the open network operating system (ONOS) controller, and the entire network is simulated using the ns-3 simulator. The simulation results demonstrate that our proposed PLD-CM scheme outperforms state-of-the-art methods in terms of various performance metrics, including the load distribution among controllers, occurrences of controller overload, round-trip time (RTT), and the average delay. Additionally, PLD-CM effectively reduces instances of controller overload by 84 %, 75 %, and 47 % compared to OpenFlow, the method proposed by Kobo et al., and the LS-CNM approach, respectively.}, archive = {J_ASOC}, author = {Vikas Tyagi and Samayveer Singh}, doi = {10.1016/j.asoc.2025.114023}, journal = {Applied Soft Computing}, month = {12}, pages = {114023}, shortjournal = {Appl. Soft. Comput.}, title = {A partial load distribution aware cluster migration approach for stable control plane in software-defined wireless sensor networks}, volume = {185}, year = {2025}, }
• (2025). Detection of alzheimer's disease using 3D-CNN model based on multimodal connectome. ASOC, 185, 114006. (www)

  @article{ , abstract = {Abnormal changes in the movement of water molecules throughout the brain related to synaptic connectivity provide information about the structure of the brain. Alzheimer's disease (AD), the most common type of dementia, is characterized by disruption of synaptic connections between neurons. This paper proposes a novel machine learning-based framework to distinguish of AD and mild cognitive impairment (MCI) that analyzes structural multimodal connectome matrix. The study consists of two stages, including preprocessing and 3D convolutional neural network (3D-CNN) model. Preprocessing pipeline includes, tractography-based probability constrained deconvolution (CSD) to generate input multimodal matrices of structural measures (FA, MD, AxD, RD). The 3D-CNN model, which trained from ADNI database with 264 individual participants, with an encoder-classifier architecture. Proposed model effectively extracts micro structural features from connectome matrix. Our framework achieved stable performance, accuracy 98.01 %, sensitivity 98.37 %, specificity 98.23 %, F1-score 97.85 %, AUC 0.99, demonstrating its potential as robust and accurate tool for early diagnosis AD.}, archive = {J_ASOC}, author = {Amir Mohammad Ahmadi Dana and Mohammad Reza Karami Mollaei and Jamal Ghasemi}, doi = {10.1016/j.asoc.2025.114006}, journal = {Applied Soft Computing}, month = {12}, pages = {114006}, shortjournal = {Appl. Soft. Comput.}, title = {Detection of alzheimer's disease using 3D-CNN model based on multimodal connectome}, volume = {185}, year = {2025}, }
• (2025). Hyperbolic spatial-temporal network for session-based recommendation. ASOC, 185, 113996. (www)

  @article{ , abstract = {The goal of Session-based recommendation is to recommend the next item a user may be interested in based on historical click data, without relying on user profiles. Compared to the data from other recommendation scenarios, session data is typically more sparse, thus Self-supervised learning (SSL), which derives ground-truth from raw data, has gradually gained attention. Existing SSL methods usually augment data by dropping or transforming original sequences in Euclidean space, leading to two problems. Firstly, in Euclidean space, it is difficult to handle the distorted distribution and hierarchical nature of session data, and secondly, the essential spatial and temporal features of session data are usually overlooked. To address these issues, we propose the Hyperbolic Spatial-Temporal Network ( HSTN ), which enhances the performance by generating positive samples with the spatial-temporal features of session data in hyperbolic space. Specifically, we first use a hyperbolic hypergraph neural network as the base encoder to mitigate the influence of distorted data distribution. Then, we employ a spatial-temporal feature learning module to generate positive samples with spatial-temporal features for contrastive learning. Extensive experiments on three real-world datasets demonstrate that our proposed method achieves improvements of 11.6 %, 17.6 %, and 38.4 %, respectively, compared to current state-of-the-art methods, under P@10 metric.}, archive = {J_ASOC}, author = {Jiaxin Zhang and Yulong Wang and Tongcun Liu and Lei Zhang and Wei Li and Jianxin Liao}, doi = {10.1016/j.asoc.2025.113996}, journal = {Applied Soft Computing}, month = {12}, pages = {113996}, shortjournal = {Appl. Soft. Comput.}, title = {Hyperbolic spatial-temporal network for session-based recommendation}, volume = {185}, year = {2025}, }
• (2025). Paradigm shift in mobility optimization with multilayered five-way decision model using rough sets. ASOC, 185, 113994. (www)

  @article{ , abstract = {To solve the complicated transportation and mobility problems that arise in Industry 4.0, this study creates a new five-way decision (FWD) architecture. It has four levels of evaluation based on rough sets (RSs) and fermatean fuzzy sets (FFSs). The new method adds two more choice areas to the usual three-way decision (TWD) framework. It establishes a more complex and varied sorting system. We begin by developing FWD specific relative gain and loss functions, establishing the foundation for the five-zone decision-making process. This study shows how reliable the threshold values are and how they can be used to set upper and lower decision limits, along with their properties. Gray relational analysis determines conditional probabilities to ensure that evaluations are correct at all levels, especially when the characteristics of the options are missing or unclear. We explore the proposed FWD model and how it can be used to find the best transportation strategy for urban mobility networks. We also address how the model can help people make decisions in a complicated, multilayered environment. We conduct tests for accuracy and reliability to verify the authenticity of the results. We can still expand the application of this approach in mobility management and overall transport planning.}, archive = {J_ASOC}, author = {Muhammad Kamran and Qingyu Zhang and Dragan Pamucar and Muhammad Tahir and Muhammad Jafar Hussain}, doi = {10.1016/j.asoc.2025.113994}, journal = {Applied Soft Computing}, month = {12}, pages = {113994}, shortjournal = {Appl. Soft. Comput.}, title = {Paradigm shift in mobility optimization with multilayered five-way decision model using rough sets}, volume = {185}, year = {2025}, }
• (2025). Intelligent compressive strength prediction of sustainable rubberised concrete using an optimised interpretable deep CNN-LSTM model with attention mechanism. ASOC, 185, 113993. (www)

  @article{ , abstract = {The increasing environmental concerns associated with waste rubber disposal, particularly from used tyres, have led to the exploration of rubberised concrete as a sustainable construction material. Rubberised concrete provides benefits like enhanced flexibility and energy absorption; however, its reduced compressive strength remains a challenge for structural applications. This study puts forward an advanced deep learning model to accurately evaluate compressive strength of rubberised concrete by combining a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) network enhanced with attention mechanism, and optimised using the enhanced firefly algorithm (EFA), featuring chaotic initialisation and nonlinear learning factor for improved convergence, for hyperparameter tuning. The proposed model introduces computing novelties: attention-guided CNN-LSTM feature fusion and chaos-enhanced firefly optimisation. Then, it is trained on an extensive dataset incorporating key mix parameters, including water, cement, supplementary cementitious materials, superplasticiser, coarse and fine aggregates, crumb and chipped rubber content, and concrete age, with validation supported by experimental tests in the laboratory. The proposed model achieves superior prediction accuracy, achieving R² values of 0.967 for training and 0.943 for testing, outperforming conventional machine learning methods. Evaluation metrics showcase the superior performance of model, with root mean square error of 2.966 MPa and 3.757 MPa for training and test data, respectively. A sensitivity analysis based on SHapley Additive exPlanations (SHAP) highlights coarse aggregate, rubber content, and concrete age as the most influential variables affecting compressive strength. By providing a highly accurate, interpretable, and cost-effective predictive tool, this research facilitates the optimisation of rubberised concrete mix design, supporting its broader adoption in sustainable construction practice.}, archive = {J_ASOC}, author = {Yang Yu and Roshan Jayathilakage and Yiyang Liu and Ailar Hajimohammadi}, doi = {10.1016/j.asoc.2025.113993}, journal = {Applied Soft Computing}, month = {12}, pages = {113993}, shortjournal = {Appl. Soft. Comput.}, title = {Intelligent compressive strength prediction of sustainable rubberised concrete using an optimised interpretable deep CNN-LSTM model with attention mechanism}, volume = {185}, year = {2025}, }
• (2025). Passenger flow prediction between subway stations based on congestion state parameters and composite neural network. ASOC, 185, 113991. (www)

  @article{ , abstract = {Traditional methods for predicting passenger flow often struggle with accuracy, data acquisition challenges, and limited applicability, failing to capture the dynamics of passenger flow variations. This study develops two novel neural network models—LSTM-RBF and LSTM-BiGRU—to predict metro station-to-station passenger flow, offering a new approach to address the above issues. Firstly, to improve the quality of acquired data, Savitzky-Golay algorithm is adopted to remove noisy data. Secondly, the kernel density estimation (KDE) is employed to convert discrete state parameters into flow curves. Thirdly, Gaussian Mixture Model (GMM) is used to classify flow curve characteristics, categorizing different station segment types. Finally, the performance of the proposed neural network models is evaluated against traditional networks using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results of the combined prediction error results under the four types of station segment scenarios show that the developed LSTM-RBF model reduces the prediction error by up to 79 %, while the developed LSTM-BiGRU model reduces the prediction error by 68 %, which indicates that the proposed methodology has different performances in different types of station segments, and effectively solves the problems of narrow scope of application of a single type of passenger flow prediction model and low prediction accuracy, low robustness of the model.}, archive = {J_ASOC}, author = {Shaofeng Wang and Guotao Gong and Yalei Liu and Feng Guo}, doi = {10.1016/j.asoc.2025.113991}, journal = {Applied Soft Computing}, month = {12}, pages = {113991}, shortjournal = {Appl. Soft. Comput.}, title = {Passenger flow prediction between subway stations based on congestion state parameters and composite neural network}, volume = {185}, year = {2025}, }
• (2025). A proximal policy optimization driven hyper-heuristic for workers constrained hybrid flow shop problem. ASOC, 185, 113990. (www)

  @article{ , abstract = {As the production environment becomes increasingly complex, the integration of soft computing techniques becomes essential for addressing resource-constrained scheduling problems. This paper delves into a worker constrained hybrid flow shop scheduling problem (WHFSP) that integrates worker resources at each processing stage. A mixed-integer linear programming (MILP) model is constructed which enables the use of mathematical solvers to obtain optimal solutions for small-scale instances. Additionally, a novel soft computing-based scheduling framework, namely a proximal policy optimization-based hyper-heuristic algorithm (PPO-HH), is proposed. It automatically selects the most suitable low-level heuristic strategies based on the current state and historical data, facilitating efficient exploration and exploitation of the complex decision space. Several low-level heuristics including perturbative and local search operators are developed to explore the solution space. Subsequently, a high-level control strategy based on proximal policy optimization is proposed. A solution quality evaluation function and a reward mechanism based on problem characteristics are formulated. This mechanism provides feedback to PPO-HH based on the degree of alignment between the actions taken by the agent and the objectives, gradually optimizing the selection of low-level heuristic strategies. Eventually, it generates a probability distribution for each low-level heuristic in the given environment. Comprehensive numerical experiments are conducted to evaluate the performance of both the MILP model and the components of the PPO-HH algorithm. The comparison results show that PPO-HH is effective and efficient for solving the WHFSP.}, archive = {J_ASOC}, author = {Shengnan Ding and Weishi Shao and Zhongshi Shao and ShengTao Peng and Dechang Pi and Jiaquan Gao}, doi = {10.1016/j.asoc.2025.113990}, journal = {Applied Soft Computing}, month = {12}, pages = {113990}, shortjournal = {Appl. Soft. Comput.}, title = {A proximal policy optimization driven hyper-heuristic for workers constrained hybrid flow shop problem}, volume = {185}, year = {2025}, }
• (2025). A self-supervised framework for mesh denoising using graph neural networks in dental applications. ASOC, 185, 113989. (www)

  @article{ , abstract = {Complete and noise-free tooth models are crucial for digital dental technology, which is widely used in dental health diagnosis. However, tooth data from real-scene scans and tooth models reconstructed by deep learning are often affected by noise, which increases the difficulty of treatment for healthcare professionals. In recent years, graph neural networks (GNNs) have made significant progress in the field of mesh denoising, but these methods often struggle to accurately capture local details when dealing with complex dental mesh models and rely on large-scale datasets requiring higher data quality. In order to obtain noise-free tooth models with realistic geometric visualization from noisy tooth data, this paper proposes a self-supervised mesh denoising framework that can automatically learn features without labels, reduce the dependence on a large amount of labeled data, and effectively deal with the challenge of data scarcity in the medical field while preserving key local details. Our core technologies include DAG-ConvNet, a graph convolutional network that combines dynamic attention mechanisms, and PosDeformNet, which fuses local and global information, to optimize normal vectors during denoising and refine vertex positions through vertex deformation networks to achieve efficient denoising of mesh data. In addition, we draw on the Total Variation Loss (TVLoss) strategy in 2D image denoising to reduce noise and geometric deformations by calculating the gradient difference of coordinates between neighboring vertices, which enhances the continuity and detail preservation of 3D meshes. We are the first to apply the self-supervised mesh denoising framework in the field of dentistry. Experimental results show that our proposed framework can effectively denoise the tooth model while preserving fine details such as texture and contour.}, archive = {J_ASOC}, author = {Yutong Hu and Yuze Shi and Haisheng Li and Li Chen}, doi = {10.1016/j.asoc.2025.113989}, journal = {Applied Soft Computing}, month = {12}, pages = {113989}, shortjournal = {Appl. Soft. Comput.}, title = {A self-supervised framework for mesh denoising using graph neural networks in dental applications}, volume = {185}, year = {2025}, }
• (2025). Beyond training horizons: A physics-informed U-net transformer for time-dependent CO₂ storage simulation in environmental applications. ASOC, 185, 113987. (www)

  @article{ , abstract = {CO 2 geological storage is a key strategy for reducing greenhouse gas emissions, requiring accurate modeling of subsurface CO 2 migration is essential for effective storage planning and risk assessment. Conventional numerical simulations, which solve time-dependent nonlinear partial differential equations, provide detailed physical insights but are computationally demanding, especially for large-scale or long-term scenarios. To improve computational efficiency, surrogate models based on machine learning have been increasingly investigated. Methods such as deep learning and physics-Informed neural networks aim to approximate the behavior of physical systems, offering potential reductions in simulation time. However, these approaches often require extensive case-specific datasets and are typically limited to fixed time horizons defined during training, which can restrict their generalizability and practical application. This study presents a time-aware surrogate modeling framework that combines convolutional neural networks with self-attention mechanisms to address these limitations. Drawing inspiration from autoregressive forecasting used in sequential learning models, the proposed approach captures temporal dependencies through iterative prediction of system states.The framework requires only a short-term numerical simulation to initialize the physical system, after which it can generate predictions for arbitrarily extended time horizons without the need for retraining. By enabling long-term forecasting, the method improves efficiency and supports repeated scenario evaluations, such as site screening and well placement optimization. Such predictive capabilities are particularly valuable in addressing environmental sustainability goals, where rapid and scalable simulations are essential for managing long-term subsurface processes under climate-related constraints.}, archive = {J_ASOC}, author = {Ye Liu and Jiahao Wang and Nan Zhang and Xiaodong Qian and Shaojun Chai}, doi = {10.1016/j.asoc.2025.113987}, journal = {Applied Soft Computing}, month = {12}, pages = {113987}, shortjournal = {Appl. Soft. Comput.}, title = {Beyond training horizons: A physics-informed U-net transformer for time-dependent CO₂ storage simulation in environmental applications}, volume = {185}, year = {2025}, }
• (2025). Evaluating CEO hubris effects on sustainable performance in the IC design industry: An integrated dynamic network DEA framework with machine learning. ASOC, 185, 113986. (www)

  @article{ , abstract = {This study introduces an integrated analytical framework combining dynamic network data envelopment analysis (DNDEA) with machine learning to assess the impact of CEO hubris on sustainable performance in the integrated circuit (IC) design industry. Our two-stage DNDEA model evaluates operational and R&D efficiency separately, incorporating intermediate factors including profit and ESG scores. We develop a novel text-based measure of CEO hubris by analyzing the contrast between confidence and conservatism language in annual shareholder reports. This hubris measure is then incorporated into predictive models, where we compare traditional linear regression against advanced machine learning approaches—support vector regression (SVR) and random forest (RF)—using cross-validation and hyperparameter optimization. The analysis reveals a significant negative correlation between CEO hubris and operational and R&D efficiency. Notably, the non-linear models (SVR and RF) demonstrate superior predictive accuracy compared to linear regression across varying levels of CEO hubris. These findings yield two primary contributions: first, they establish the critical need for monitoring hubristic leadership behavior in innovation-intensive industries, given their detrimental effect on organizational efficiency. Second, they validate the effectiveness of combining text analytics, DNDEA efficiency metrics, and machine learning for evaluating leadership impact on firm performance. This methodology provides a comprehensive framework for analyzing leadership dynamics in the IC design sector and offers an adaptable template for similar analyses across technology-driven industries.}, archive = {J_ASOC}, author = {Sheng-Wei Lin and Yu-Rou Lin}, doi = {10.1016/j.asoc.2025.113986}, journal = {Applied Soft Computing}, month = {12}, pages = {113986}, shortjournal = {Appl. Soft. Comput.}, title = {Evaluating CEO hubris effects on sustainable performance in the IC design industry: An integrated dynamic network DEA framework with machine learning}, volume = {185}, year = {2025}, }
• (2025). FSAMLM: A few-shot adaptation multimodal large model for cross-domain fault diagnosis. ASOC, 185, 113985. (www)

  @article{ , abstract = {Recent advances in deep learning have demonstrated remarkable performance in mechanical fault diagnosis. However, most existing approaches are designed for a single data modality or a single task, limiting their flexibility and generalization ability. Moreover, common knowledge across different tasks remains largely unexploited. Inspired by the success of large-scale models in natural language processing and computer vision, integrating multimodal data with multi-task learning strategies may significantly improve the performance of fault diagnosis models. Nonetheless, a significant gap exists between general pre-trained knowledge and domain-specific expertise, posing considerable challenges for effective integration. To address these issues, we propose a novel few-shot adaptation multimodal large model (FSAMLM) that efficiently captures shared representations through a parameter-efficient fine-tuning strategy, structured in two stages. Specifically, we first design a low-rank adaptation meta-learning (LoRAML) framework, which employs low-rank decomposition on pre-trained parameters to reduce computational complexity and improve robustness. This approach not only accelerates adaptation to new few-shot tasks but also preserves pre-training knowledge effectively. During the second fine-tuning stage, we implement target-domain training with limited samples and introduce a training-free inference option for real-world deployment. Experimental validation using six public datasets demonstrates FSAMLM’s superior domain generalization capability in few-shot fault diagnosis tasks compared to existing methods. The code repository is publicly available at: https://github.com/ohhyeeaah/FSAMLM-for-fault-diagnosis .}, archive = {J_ASOC}, author = {Xin Zhang and Shixi Liu and Li Jiang and Yibing Li}, doi = {10.1016/j.asoc.2025.113985}, journal = {Applied Soft Computing}, month = {12}, pages = {113985}, shortjournal = {Appl. Soft. Comput.}, title = {FSAMLM: A few-shot adaptation multimodal large model for cross-domain fault diagnosis}, volume = {185}, year = {2025}, }
• (2025). Simplified directional message passing neural network for prediction of protein-ligand binding affinity. ASOC, 185, 113983. (www)

  @article{ , abstract = {Effective prediction of protein-ligand binding affinity can dramatically accelerate drug discovery. Recent studies have shown that graph neural networks have great potential in the field of binding affinity prediction. However, previous models for binding affinity prediction have not made good use of the three-dimensional (3D) information in complex graphs, and the presence of redundancy during the message passing process has adversely impacted model performance. In light of these limitations, we propose a novel binding affinity prediction model named Simplified Directional Message Passing Neural Network (SimdNet). Our network comprises a message passing module that capitalizes on the quantum information of the protein-ligand system, encompassing both radial and angular information, while mitigating redundancy during message passing. Additionally, we introduce a readout module that focuses on local interactions. In a more realistic and challenging evaluation scenario—using a larger dataset with mixed-quality data—our model outperforms the strongest baseline with a 3.4 % lower RMSE, and through ablation experiments, we validate the efficacy of each module. Our implementation is available at https://github.com/AILBC/SimdNet .}, archive = {J_ASOC}, author = {Yankai Yu and Yongquan Jiang and Yan Yang and Zhuyang Xie and Qun Lu}, doi = {10.1016/j.asoc.2025.113983}, journal = {Applied Soft Computing}, month = {12}, pages = {113983}, shortjournal = {Appl. Soft. Comput.}, title = {Simplified directional message passing neural network for prediction of protein-ligand binding affinity}, volume = {185}, year = {2025}, }
• (2025). A YOLOv3-powered edge computing technique for real-time rear-end collision prediction in autonomous vehicles. ASOC, 185, 113981. (www)

  @article{ , abstract = {Rear-end collisions are the prime cause of traffic casualties and to avoid these casualties, rear-end collision prediction has procured recognition for smart transportation. The traditional research on rear-ends collision prediction techniques involved Perception-Reaction Time (PRT) parameter which leads to reduced adaptive competence for different PRTs and yields degraded performance. Thus, learning-based techniques were proposed to deal with the related issues and are facing challenges with reference to feature extraction and prediction performance. The paper presents an optimized YOLOv3 based Rear-end Collision Prediction (YOLO-RECP) to address the traffic casualties caused by rear-end collisions. In YOLO-RECP, the real-time NGSIM dataset is streamlined and expanded to extenuate the class imbalance issue using Smote Gaussian Mixture (SGM) technique. After pre-processing the dataset, it acts as input for testing and training of the optimized YOLOv3 network. Then, the proposed system model outputs different classes of collision warnings in real-time. The simulation results demonstrate that YOLO-RECP outperformed baseline YOLOv3, RCPM, MCWA, Honda, and Berkeley techniques by predicting the possible rear-end collisions.}, archive = {J_ASOC}, author = {Divya Punia and Rajender Kumar}, doi = {10.1016/j.asoc.2025.113981}, journal = {Applied Soft Computing}, month = {12}, pages = {113981}, shortjournal = {Appl. Soft. Comput.}, title = {A YOLOv3-powered edge computing technique for real-time rear-end collision prediction in autonomous vehicles}, volume = {185}, year = {2025}, }
• (2025). Machine learning-based adaptive large neighborhood search algorithm for the integrated vessel scheduling and speed optimization problem in the compound channel. ASOC, 185, 113980. (www)

  @article{ , abstract = {The carbon emission from vessel navigating in the channel accounts for about 61 % of the total emissions in port areas. Considering an effective means of reducing emissions, namely, speed adjustment, this study deals with an integrated problem of vessel scheduling and speed optimization (VSSOP) in the channel. This study considers the complex structure of a compound channel, i.e., containing both one-way and two-way lanes with different navigation rules. We also focus on the effects of meteorological conditions (winds, waves and currents) on the vessel stall, and tidal restrictions on the time window for large vessels to pass through the channel. Thus, a mixed integer programming (MIP) model for the VSSOP is proposed to control the carbon emissions in the channel. Then, we develop a machine learning-based adaptive large neighborhood search (ALNS) approach, where the ALNS is used to solve the proposed MIP in real cases and the dynamic machine learning approach helps to evaluate and fit the complex effects of multiple meteorological conditions on the vessel sailing speed. The dynamic parallel mechanism is further introduced to improve the fitting accuracy of the machine learning part without increasing the running time of the ALNS. The experimental results reveal that the machine learning-based ALNS approach can be applied in practice. Additionally, valuable managerial insights for port operators are obtained to aid in vessel traffic management.}, archive = {J_ASOC}, author = {Jian Du and Shan Lin and Liming Guo and Jianfeng Zheng}, doi = {10.1016/j.asoc.2025.113980}, journal = {Applied Soft Computing}, month = {12}, pages = {113980}, shortjournal = {Appl. Soft. Comput.}, title = {Machine learning-based adaptive large neighborhood search algorithm for the integrated vessel scheduling and speed optimization problem in the compound channel}, volume = {185}, year = {2025}, }
• (2025). Improving mutation strategies in differential evolution with a new pbest selection mechanism. ASOC, 185, 113978. (www)

  @article{ , abstract = {Differential evolution, which belongs to a group of population-based algorithms, has received a lot of research attention since its introduction in 1995. A population-based algorithm is required to guide individuals to visit potentially better basins of attraction in the search space when searching for a globally optimal solution. Additionally, individuals need to interact with each other during an evolutionary process to explore the search space effectively. In this paper, we propose a novel pbest selection mechanism for DE/current-to-pbest mutation strategy and its variants designed to enhance the potential for exploration of different attraction basins. The proposed mechanism enforces a minimal distance between the selected pbest individual and all other better individuals. This means that possible candidates for the pbest individual, used in mutation, are further spaced apart. As a result, the likelihood that the new trial vector will be generated in a different attraction basin of the search space is increased. The mechanism is incorporated into the L-SHADE, jSO, and L-SRTDE algorithms, and its effectiveness is evaluated using CEC’24 benchmark functions. Experimental results demonstrate improvements in the performance of the selected algorithms, particularly in higher-dimensional problem instances.}, archive = {J_ASOC}, author = {Jan Popič and Borko Bošković and Janez Brest}, doi = {10.1016/j.asoc.2025.113978}, journal = {Applied Soft Computing}, month = {12}, pages = {113978}, shortjournal = {Appl. Soft. Comput.}, title = {Improving mutation strategies in differential evolution with a new pbest selection mechanism}, volume = {185}, year = {2025}, }
• (2025). Transformer-guided optimization of railway alignments using an adaptive rapidly-exploring random tree algorithm. ASOC, 185, 113977. (www)

  @article{ , abstract = {Railway alignment design is a crucial part of a railway project. Despite the widespread success of computer-aided alignment optimization methods in determining alignments, effectively exploring the objective function’s descent direction (OFDD) remains challenging, particularly when navigating complex alignment search spaces. To address this issue, it is essential to comprehensively consider key factors, including the global and local environment, explored and unexplored search spaces, as well as established alignment search strategies and potential new ones that may emerge during the OFDD optimizing process. Therefore, an alignment-oriented Transformer framework is formulated in this work. In this framework, various real-world railway cases are input into a stacked Transformer framework to learn an optimized OFDD strategy. Specifically, the model handles regular inputs (i.e., global and local contexts, long-term goals) and irregular inputs (i.e., historical paths) using two separate stacked Transformer encoders. Afterward, an adaptive rapidly-exploring random tree star (Ada-RRT-star) method is developed by integrating the Transformer framework’s output to guide RRT’s search direction as well as to enhance the solution quality. Ultimately, the proposed method is applied to a realistic railway case, where the results demonstrate its superiority over the conventional 3D-RRT-star algorithm in terms of solution quality. Besides, the best alignment generated by the Ada-RRT-star also outperforms the manually-designed alignment.}, archive = {J_ASOC}, author = {Xinjie Wan and Hao Pu and Paul Schonfeld and Yang Ran and Taoran Song and Lihui Peng}, doi = {10.1016/j.asoc.2025.113977}, journal = {Applied Soft Computing}, month = {12}, pages = {113977}, shortjournal = {Appl. Soft. Comput.}, title = {Transformer-guided optimization of railway alignments using an adaptive rapidly-exploring random tree algorithm}, volume = {185}, year = {2025}, }
• (2025). A label enhancement based positive-unlabeled hybrid network for pump bearing intelligent fault diagnosis. ASOC, 185, 113976. (www)

  @article{ , abstract = {Bearings are important support parts for rotating machinery such as pumps, and the application of machine learning algorithms has brought the fault diagnosis of bearings to a more intelligent stage. However, with the scarcity of target fault data and the lack of accurate labeling for critical data, commonly used data-driven fault diagnosis methods had its limitations. Inspired by semi-supervised learning, hypergraph and knowledge distillation theories, a hybrid PUHGNN network based on label augmentation was proposed in this paper. Firstly, a hypergraph neural network (HGNN) structure based on the multi-resolution signal was proposed to measure the correlation at multiple scales to make difference and connections between different labels. Secondly, the HGNN network is improved by combining HGNN and Positive-Unlabeled (PU) Learning ideas to form a new PUHGNN label enhancing mechanism which will solve the lacking of labels. Lastly, a soft-label-based label selection method is proposed to dynamically judge the similarity of samples to reiterate, which will make the otherwise indistinguishable faults more explicit. In experimental session, the CWRU dataset and the enviormental protection pump bearing datasets were applied to conduct unbalance, mislabel, extreme mislabel and ablation experiments. The result shows that the label enhancement is not only necessary but significant in the unbalanced under-labeled datasets, furthermore, the PUHGNN has more obvious enhancement compared to other methods.}, archive = {J_ASOC}, author = {Jiaxing zhu and Junlan Hu and Buyun Sheng}, doi = {10.1016/j.asoc.2025.113976}, journal = {Applied Soft Computing}, month = {12}, pages = {113976}, shortjournal = {Appl. Soft. Comput.}, title = {A label enhancement based positive-unlabeled hybrid network for pump bearing intelligent fault diagnosis}, volume = {185}, year = {2025}, }
• (2025). Rural road extraction from remote sensing images based on multi-view contextual information and multi-stage features. ASOC, 185, 113975. (www)

  @article{ , abstract = {Accurate extraction of rural roads from high-resolution optical remote sensing images is of great significance to the development of rural areas, road navigation, rural land resource planning and other applications. Different from urban roads, rural ones with complex terrain backgrounds are often slender and winding, thereby making them more susceptible to vegetation cover. In order to improve the reliability and accuracy of rural road extraction, a Complex Rural Road Extraction Network (CRRENet) is proposed in this work, which consists of five parts: feature encoder, Multi-view Contextual Information Extraction Module (MCIEM), Multi-stage Feature Fusion Module (MFFM), Channel Coordinate Attention Mechanism (CCAM) and feature decoder. The MCIEM extracts the multi-view contextual information by the parallel dilated convolution with different dilation rates. To avoid the loss of image details, the MFFM integrates different feature maps from the downsampling stages. By adjusting the weights of feature maps, the CCAM enables the network to self-adaptively suppress the background noise and focus on the road foreground. Ablation and comparison validate CRRENet's superiority.}, archive = {J_ASOC}, author = {Langping Li and Jizheng Yi and Pengyu Lei and Hengkai Lou and Xiaoyao Li and Hui Lin}, doi = {10.1016/j.asoc.2025.113975}, journal = {Applied Soft Computing}, month = {12}, pages = {113975}, shortjournal = {Appl. Soft. Comput.}, title = {Rural road extraction from remote sensing images based on multi-view contextual information and multi-stage features}, volume = {185}, year = {2025}, }
• (2025). Joint stage features modulation for progressive blind face restoration. ASOC, 185, 113974. (www)

  @article{ , abstract = {A significant challenge for Blind Face Restoration (BFR) is to cope with the degraded information of unknown parameters in face images. The BFR method has evolved from non-prior to prior-based methods, but there are still some shortcomings. The quality of priors seriously affects the restoration results, especially in scenarios with severe degradation. Simultaneously encoding or modulating degraded images directly into the restoration process can introduce degraded information, leading to poor visual perception. Therefore, we propose a progressive restoration model with the joint stage features modulation, named JSFM-GAN. JSFM-GAN can be seen as having two stages. In the first stage, the LQ image is modulated with the facial resolution map to provide a rough structure for recovery. In the second stage, Joint Stage Feature Modulation (JSFM) utilizes the LQ images and stage features for joint modulation on multiple scales to balance fidelity and realism by combining clean spatial information of stage features and the tonal structure of LQ images. At the same time, the Up-Sampling Feature Supplement Block (UFSB) is used to reduce information loss due to channel fading and improve the network’s focus on face components and textures. In addition, we use the stage reconstruction loss and adjusted facial parsing maps to enhance the realism and symmetry of the generated results. Experiments with JSFM-GAN on synthetic and real-world datasets achieve good results, demonstrating the superior performance of our method.}, archive = {J_ASOC}, author = {Xianlun Tang and Xiaodong Qian and Jie Li and Binyu Lu and Wuquan Deng and Weisheng Li}, doi = {10.1016/j.asoc.2025.113974}, journal = {Applied Soft Computing}, month = {12}, pages = {113974}, shortjournal = {Appl. Soft. Comput.}, title = {Joint stage features modulation for progressive blind face restoration}, volume = {185}, year = {2025}, }
• (2025). Mamba-based multi-branch cost aggregation for stereo matching. ASOC, 185, 113973. (www)

  @article{ , abstract = {This study presents Mamba-Based Multi-Branch Cost Aggregation for Stereo Matching (MMBStereo), an innovative real-time stereo matching framework with high performance. The core innovation lies in the Mamba-based multi-branch cost aggregation network, which uses a unique three-branch aggregation strategy. The Mamba Aggregation Branch integrates the State Space Model from the Mamba structure, replacing conventional convolution and Transformer methods, significantly enhancing network performance and efficiency. The Spatial Aggregation Branch addresses the loss of spatial texture information, improving the scene’s contextual representation. Meanwhile, the Edge Aggregation Branch enhances edge responses, improving object boundary detection accuracy. Through a carefully designed multi-branch fusion strategy, the framework improves disparity prediction accuracy while maintaining real-time inference. Our method achieves competitive accuracy with non-real-time stereo matching frameworks, surpassing existing lightweight solutions in the widely recognized KITTI benchmark tests.}, archive = {J_ASOC}, author = {Xingyuan Lu and Yanbing Xue and Leida Li and Shiyin Li and Zan Gao}, doi = {10.1016/j.asoc.2025.113973}, journal = {Applied Soft Computing}, month = {12}, pages = {113973}, shortjournal = {Appl. Soft. Comput.}, title = {Mamba-based multi-branch cost aggregation for stereo matching}, volume = {185}, year = {2025}, }
• (2025). Multilinear hyperspectral unmixing based on autoencoder and recurrent neural network. ASOC, 185, 113972. (www)

  @article{ , abstract = {Spectral unmixing techniques estimate the endmember spectra and corresponding abundance fractions that constitute the pixels of hyperspectral remote sensing images, revealing the mixing mechanisms of materials within pixels. In recent years, deep learning has shown significant potential in advancing spectral unmixing, particularly in nonlinear scenarios. However, most existing nonlinear models rely on bilinear mixing frameworks, with limited focus on high-order nonlinear models. This restricts their ability to capture complex interactions such as multiple light scattering events. To address this issue, this work proposes an unsupervised unmixing method leveraging an autoencoder network framework and the multilinear mixing model (MLM). It employs a recurrent neural network (RNN) in the decoder to simulate the multiple scattering of light between materials. Unlike conventional multilinear approaches that rely on explicit mathematical formulations, the proposed method leverages the RNN to automatically learn and approximate the nonlinear interactions of light. Moreover, the RNN weights are adaptively updated during training and interpreted as transition probabilities representing further light interactions among materials, endowing the model structure with explicit physical interpretation. Besides, a new stopping criterion is also designed, which ensures better RNN weights are obtained during backpropagation. Experiments conducted on both synthetic and real datasets demonstrate the better performance of the proposed method.}, archive = {J_ASOC}, author = {Zehui Jin and Xiaorui Yi and Yue Liu and Hongjuan Zhang}, doi = {10.1016/j.asoc.2025.113972}, journal = {Applied Soft Computing}, month = {12}, pages = {113972}, shortjournal = {Appl. Soft. Comput.}, title = {Multilinear hyperspectral unmixing based on autoencoder and recurrent neural network}, volume = {185}, year = {2025}, }
• (2025). Deadline-constrained and bi-objective workflow scheduling with fuzziness in cloud computing systems. ASOC, 185, 113970. (www)

  @article{ , abstract = {Deploying intelligent expert systems on the cloud is a cost-effective solution for handling compute-intensive service requests, where each request is treated as a workflow processing procedure. Cloud workflows break down complex service requests into smaller tasks and take advantage of automatic maintenance services provided by the cloud workflow management system (CWMS). CWMS can orchestrate task execution, handle dependencies, and direct dynamic scaling of resources based on workload demands. For a CWMS, the main challenge lies in the uncertainty of workflow scheduling, i.e., the processing time, the data transmission time, and the due date are not crisp values. This paper investigates the problem of bi-objective workflow scheduling under fuzziness, aiming to minimize both the total rental cost and the degree of user dissatisfaction. Triangular fuzzy numbers are used to represent the uncertainty of temporal parameters. Two general pricing models in cloud systems are considered: on-demand and reserved price structures. A bi-objective fuzzy workflow scheduling framework is proposed, which consists of workflow sequencing, fuzzy solution generation and solution improvement components. The workflow sequencing component determines the priorities of the workflows. The fuzzy solution generation component assigns tasks to appropriate resources. The Simulated Annealing with Variable Neighborhood Search (SAVNS) method is developed for the solution improvement component. The experimental results demonstrate that the proposal can achieve better effectiveness and robust performance than the baseline algorithms compared. The proposed method can offer practical solutions for CWMS to optimize cost-performance trade-offs in uncertain environments.}, archive = {J_ASOC}, author = {Jie Zhu and Fengmei Liu and Jingzhe Sun and Haiping Huang}, doi = {10.1016/j.asoc.2025.113970}, journal = {Applied Soft Computing}, month = {12}, pages = {113970}, shortjournal = {Appl. Soft. Comput.}, title = {Deadline-constrained and bi-objective workflow scheduling with fuzziness in cloud computing systems}, volume = {185}, year = {2025}, }
• (2025). Detail-aware semantic segmentation network for brain tumor MRI images combining multi-frequency directional filtering and lifting wavelets. ASOC, 185, 113969. (www)

  @article{ , abstract = {Brain tumors segmentation in Magnetic Resonance Imaging (MRI) images poses significant challenges owing to the uncertain location and size of the tumors, the difficulty in describing their boundaries, and the fuzzy demarcation of diseased tissues. Although U-Net and its recent variants have emerged as leading models for semantic segmentation in medical imaging, they still face structural limitations. These limitations cause the erosion of detail information during downsampling and poor performance in segmenting small lesions when handling targets of varying sizes, indicating a lack of detail handling capability. To counteract these issues, we designed a segmentation model that enhances detail features using frequency information. To reduce the loss of feature information during downsampling, we developed a downsampling module based on lifting wavelets. By lifting wavelets to group and integrate features according to frequency from high to low, we reduce feature resolution while enhancing information transmission and minimizing feature information loss. In our designed multi-frequency directional filtering edge feature extraction module, we extract low-frequency and high-frequency features and construct a dual-channel multi-directional filtering combination. This combination extracts directional information from low-frequency and high-frequency features separately, increasing the multi-angle directional information of the features and enriching the detailed information such as direction and position within the features. On the BraTS2018, BraTS2020, and BraTS2024 brain tumor datasets, our model demonstrated optimal results compared to 14 other advanced models. The average Dice Similarity Coefficients are 78.48 %, 79.80 %, and 74.35 %, while the 95th percentile Hausdorff Distances are 5.75, 6.60, and 7.72. Our code link is https://github.com/Eric-H8/BraTS_Seg_Model .}, archive = {J_ASOC}, author = {Xin Hua and Zhijiang Du and Hongjian Yu and Zibo Li and Qiaohui Lu and Hui Zhao}, doi = {10.1016/j.asoc.2025.113969}, journal = {Applied Soft Computing}, month = {12}, pages = {113969}, shortjournal = {Appl. Soft. Comput.}, title = {Detail-aware semantic segmentation network for brain tumor MRI images combining multi-frequency directional filtering and lifting wavelets}, volume = {185}, year = {2025}, }
• (2025). Extracting knowledge from limited data: An updated review of data-driven and model-driven few-shot learning for agriculture. ASOC, 185, 113968. (www)

  @article{ , abstract = {Deep learning has demonstrated considerable success in agricultural applications. However, its conventional implementations heavily depend on large-scale labelled datasets—a requirement that is often impractical in agriculture due to data scarcity, high annotation costs, or environmental variability. While insufficient training data can significantly limit the performance of standard deep learning models, Few-Shot Learning (FSL) has emerged as a transformative paradigm, enabling robust model training with minimal labelled samples by utilising limited data for training instead. Despite its potential, a critical review assessing how FSL addresses expert system challenges in agriculture remains notably absent. This paper attempts to fill this void by presenting an updated comprehensive review of FSL's applications in agriculture. We categorise FSL methodologies into two primary approaches: data processing-driven and model learning-driven. Data processing–driven approaches address data scarcity by enriching representational diversity through synthetic samples generated with models such as generative adversarial networks, or by transferring knowledge from related domains to improve generalisation. In contrast, model learning–driven strategies confront the same challenge through specialised architectures and optimisation techniques that enable effective generalisation from limited samples. Within this taxonomy, data processing–driven paradigms include transfer learning and generative artificial intelligence, while model learning–driven paradigms cover metric learning methods such as Siamese or prototypical networks, together with model-based and optimisation approaches designed for efficient generalisation. Our analysis pinpoints cutting-edge technologies within each sector, shedding light on overlooked areas and opportunities where FSL can harness limited data to yield promising outcomes when used to solve problems in agriculture.}, archive = {J_ASOC}, author = {Kam Meng Goh and Usman Ullah Sheikh and Jun Kit Chaw and Weng Kin Lai and Weng Chun Tan and Santhi Krishnamoorthy}, doi = {10.1016/j.asoc.2025.113968}, journal = {Applied Soft Computing}, month = {12}, pages = {113968}, shortjournal = {Appl. Soft. Comput.}, title = {Extracting knowledge from limited data: An updated review of data-driven and model-driven few-shot learning for agriculture}, volume = {185}, year = {2025}, }
• (2025). Rapid attribute and scale selection with adaptive three-way sampling and neighborhood rough set. ASOC, 185, 113966. (www)

  @article{ , abstract = {As a key issue in knowledge reduction for multi-scale data, attribute and scale selection has attracted increasing attention in recent years. However, with the rapid growth of data volumes, most existing methods are inefficient for large-scale multi-scale data and struggle to effectively handle heterogeneous multi-scale data, significantly limiting their practical applications. To address this situation, this paper proposes a rapid attribute and scale selection method to deal with large nominal-and-numerical mixed multi-scale decision systems (NN-MDSs). First, the theory and algorithm of adaptive incremental support vector data description (AISVDD) are presented. The AISVDD algorithm overcomes the limitations of traditional support vector data description methods and efficiently processes both class-balanced and class-imbalanced data. By using the algorithm, support vectors can be quickly obtained from large data. Next, an adaptive three-way sampling technique is derived by combining AISVDD and three-way decision. With this technique, support vectors are extracted as sampling results and put into the boundary region, and outliers are seen as noise and put into the negative region. This significantly reduces the data size and improves the data quality. Then, a neighborhood rough set model is built to describe NN-MDSs. Multiple concepts and properties are discussed in the model. Finally, a heuristic attribute and scale selection algorithm is designed to simultaneously choose attributes and scales from the sampled NN-MDS. Detailed experiments demonstrate the effectiveness and superiority of the proposed method. The method performs better than state-of-the-art attribute and scale selection methods on both computational efficiency and classification performance under six benchmark classifiers. It is powerful in handling large NN-MDSs with complex characteristics. This work provides new insights into the complex multi-scale data processing.}, archive = {J_ASOC}, author = {Shujiao Liao and Nan Zhou and Ling Wei and Weiping Ding and Yidong Lin}, doi = {10.1016/j.asoc.2025.113966}, journal = {Applied Soft Computing}, month = {12}, pages = {113966}, shortjournal = {Appl. Soft. Comput.}, title = {Rapid attribute and scale selection with adaptive three-way sampling and neighborhood rough set}, volume = {185}, year = {2025}, }
• (2025). A hybrid wind-PV power generation forecasting model based on uncertainty quantification. ASOC, 185, 113965. (www)

  @article{ , abstract = {Accurate forecasts of wind and photovoltaic power generation are essential for the stable operation of power systems. However, traditional models struggle to quantify the uncertainty during the forecasting process. This paper proposes a hybrid forecasting model to quantify uncertainty. For low shared information, uncertainty is quantified separately and forecasted using independent models; for high shared information, a multi-task model enables joint forecasting. Hyperparameters are adjusted with an evolutionary optimization algorithm. Experimental results show that in short-term forecasting, the model achieves the best performance, with RMSE of 0.0362 and R 2 of 0.9561 for PV, and RMSE of 0.0572 and R 2 of 0.9384 for wind during October to December. The directional error metric indicates superior performance of the proposed model in wind power forecasting. In long-term and comprehensive forecasting experiments, the Diebold-Mariano test confirms the statistical superiority of the proposed model over traditional models. The proposed approach provides an effective model to support real-world decision-making in renewable energy integration.}, archive = {J_ASOC}, author = {Xuan Yang and Yunxuan Dong and Lina Yang and Thomas Wu}, doi = {10.1016/j.asoc.2025.113965}, journal = {Applied Soft Computing}, month = {12}, pages = {113965}, shortjournal = {Appl. Soft. Comput.}, title = {A hybrid wind-PV power generation forecasting model based on uncertainty quantification}, volume = {185}, year = {2025}, }
• (2025). Coverage exploration of unknown obstacle-cluttered environments using a swarm of ground robots. ASOC, 185, 113964. (www)

  @article{ , abstract = {This paper introduces a coverage exploration algorithm for unknown obstacle-cluttered environments using a swarm of ground robots. A key contribution of this work is the proposed fitness function, which balances multiple exploration objectives and encourages robots to disperse effectively, avoiding excessive overlapping visits. The robots are assumed to start from a single corner of the environment, reflecting practical situations where pre-distributing them is not feasible. This setup highlights a key feature of the algorithm, as it enables self-organization and effective distribution of the robots throughout the environment. The robustness of the method is demonstrated through experiments in various environmental setups, showing its resilience to different obstacle structures and reliable performance across diverse scenarios. The approach also leverages the benefits of swarm behavior, where an increasing number of robots improves exploration efficiency through enhanced collaboration and coverage. The algorithm is evaluated against a swarm random walk approach and two multi-robot meta-heuristic methods, significantly outperforming them in terms of coverage efficiency and robustness.}, archive = {J_ASOC}, author = {Khalil Al-rahman Youssefi and Wilfried Elmenreich}, doi = {10.1016/j.asoc.2025.113964}, journal = {Applied Soft Computing}, month = {12}, pages = {113964}, shortjournal = {Appl. Soft. Comput.}, title = {Coverage exploration of unknown obstacle-cluttered environments using a swarm of ground robots}, volume = {185}, year = {2025}, }
• (2025). Federated deep embedded clustering under privacy protection. ASOC, 185, 113963. (www)

  @article{ , abstract = {Deep clustering, a prominent research focus in data mining, utilizes deeply embedded features to reveal the intrinsic statistical structure of data. However, existing deep clustering models rely on centralized datasets, which are impractical in scenarios with data silos and privacy constraints, leading to degraded clustering performance. Considering the privacy protection characteristics of federated learning, this paper incorporates the idea of federated learning into deep clustering, and proposes a federated deep embedding clustering (FDEC) model under privacy protection. FDEC follows a universal client-server structure, coordinating training between clients through a central server to obtain a unified central model. The server updates the global deep embedding and cluster centers based on a hybrid federated averaging strategy, while each client conducts two-stage deep clustering on local data without sharing raw data. To enhance robustness under non-independent and identically distributed (non-IID) conditions, the hybrid strategy improves parameter aggregation effectiveness. Experimental results on both IID and non-IID datasets demonstrate that FDEC is more robust and consistently outperforms centralized deep clustering methods.}, archive = {J_ASOC}, author = {Xiao Xu and Hong Liao and Xu Yang}, doi = {10.1016/j.asoc.2025.113963}, journal = {Applied Soft Computing}, month = {12}, pages = {113963}, shortjournal = {Appl. Soft. Comput.}, title = {Federated deep embedded clustering under privacy protection}, volume = {185}, year = {2025}, }
• (2025). WamGLM: A multimodal large-scale language model for wafer map defect information in-depth query through multi-turn dialogue based on prototypical supervised contrastive learning. ASOC, 185, 113962. (www)

  @article{ , abstract = {To ensure production efficiency and process stability in semiconductor manufacturing, it is of critical importance to detect wafer map defects and perform information query for tracing and solving problems during the manufacturing process. Numerous vision models based on deep learning have been successfully applied to wafer map defect recognition (WMDR), yielding remarkable results. However, the dynamic and in-depth querying of wafer map defect information remains relatively underexplored. Leveraging the rapid advancements in multimodal large language models (MLLMs), this paper proposes a novel approach for wafer map defect information query (WMDIQ). First, following the paradigm of employing cross-modal alignment model to bridge vision and language models, an end-to-end response MLLM: general language model for wafer map (WamGLM), is constructed for WMDIQ. Concurrently, by designing an interactive dialogue framework between large language models (LLMs), the first large-scale multi-turn dialogue dataset: visual multi-turn question answering dataset for wafer map defects (WaferMapVMQA Dataset), is constructed for wafer map defect analysis. Subsequently, WamGLM is trained using a two-stage fine-tuning strategy. In the first stage, a visual fine-tuning method based on prototypical supervised contrastive learning (PSCL) is introduced to enhance the intra-class compactness and inter-class separability of defect features. In the second stage, language fine-tuning is conducted using the WaferMapVMQA Dataset to infuse specialized knowledge into WamGLM. To validate the effectiveness and superiority of the proposed method, experiments are conducted on a real wafer map dataset. The results demonstrate that the proposed method significantly outperforms other methods in both defect recognition performance and information query response performance. Our code is available at: https://github.com/ZihaoLei/WamGLM .}, archive = {J_ASOC}, author = {Shulong Gu and Zihao Lei and Guangrui Wen and Quanning Xu and Zhaojun Steven Li and Xuefeng Chen and Chunsheng Yang}, doi = {10.1016/j.asoc.2025.113962}, journal = {Applied Soft Computing}, month = {12}, pages = {113962}, shortjournal = {Appl. Soft. Comput.}, title = {WamGLM: A multimodal large-scale language model for wafer map defect information in-depth query through multi-turn dialogue based on prototypical supervised contrastive learning}, volume = {185}, year = {2025}, }
• (2025). Representing online reviews using interval type-2 fuzzy Z-numbers for ranking energy-saving appliances. ASOC, 185, 113961. (www)

  @article{ , abstract = {As e-commerce develops and the green consumption concept gains popularity, consumers are increasingly inclined to purchase energy-saving home appliances through e-commerce platforms. However, they often face technical complexities related to specialized energy-saving attributes and an overwhelming number of online reviews. To address these challenges, we have integrated energy-saving features into our online review analysis, incorporating weight calculations. Notably, we propose and prove a method that transforms online review information into interval type-2 fuzzy Z-numbers (IT2FZNs), which comprehensively represent the information to support product ranking. First, based on the energy-saving attributes of energy-saving appliances and their online review data, we extract energy-saving features and online review features, respectively. We then use a combination of TF-IDF-based text mining and BWM-based expert evaluation to determine the weight of each feature. Next, we convert the energy-saving feature data into IT2FZNs according to specific rules. The online review feature data is converted into interval type-2 fuzzy sets (IT2Fs) by considering the sentiment classification results and the accuracy and robustness of the model, and is further combined with the reliability of online reviews to form IT2FZNs. Finally, the alternative products are ranked based on the constructed decision matrix, and the final ranking results are determined. The method's effectiveness and practicality have been demonstrated using real data from energy-saving refrigerators on the JingDong (JD.com) platform, and its robustness and superiority have been further substantiated through comparative experiments.}, archive = {J_ASOC}, author = {Yue Xiao and Ming Li and Ying Li and Hongde Liu}, doi = {10.1016/j.asoc.2025.113961}, journal = {Applied Soft Computing}, month = {12}, pages = {113961}, shortjournal = {Appl. Soft. Comput.}, title = {Representing online reviews using interval type-2 fuzzy Z-numbers for ranking energy-saving appliances}, volume = {185}, year = {2025}, }
• (2025). Multi-criteria linguistic optimization for covert communication in secure LLM-based steganography. ASOC, 185, 113960. (www)

  @article{ , abstract = {This paper presents a novel framework for covert communication through secure steganography using large language models (LLMs). Our approach leverages multi-criteria linguistic optimization to encode secret information directly into stylistic features of auto-regressively generated text. This strategy balances embedding capacity with naturalness and coherence. The secret message is partitioned into fixed-size blocks. Each block is embedded into binary stylistic feature vectors via a surjective linear mapping, which introduces redundancy. This redundancy enables the use of a history-aware cost function that selects stylistic vectors to minimize abrupt transitions and preserve fluency across sentences. Candidate sentences are generated by prompting LLMs with contextual and stylistic constraints. Rejection sampling then ensures exact feature matching and high linguistic quality. Experimental evaluation in multiple LLMs, diverse text contexts, and parameter settings demonstrates effective embedding capacities of up to 0.30 bits per token while maintaining strong linguistic naturalness, validated through perplexity, lexical diversity, readability, and a linguistic acceptability metric. Importantly, decoding recovers the full secret with zero error under ideal conditions. This confirms the reliability of the method. The current work focuses on embedding efficiency and imperceptibility. Robustness against active text alterations and formal undetectability assessments remain open challenges for future research. The proposed multi-criteria linguistic optimization framework offers a promising avenue for advanced covert communication by harmonizing secure information embedding with fluent, human-like language generation.}, archive = {J_ASOC}, author = {Kamil Woźniak and Marek R. Ogiela and Lidia Ogiela}, doi = {10.1016/j.asoc.2025.113960}, journal = {Applied Soft Computing}, month = {12}, pages = {113960}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-criteria linguistic optimization for covert communication in secure LLM-based steganography}, volume = {185}, year = {2025}, }
• (2025). Cooperative driving route planning based on asymmetric multi-agent path planning problem with limited service area constraints. ASOC, 185, 113959. (www)

  @article{ , abstract = {Mobility and logistics service providers operating a vehicle fleet and serving several targets with each vehicle are faced with the challenge of planning efficient and cooperative driving routes while considering all the vehicles. To address this challenge, we present a novel driving route planning method based on an Asymmetric Multi-Agent Path Planning (AMAPP) problem, a variation of the classical Multiple Traveling Salesmen Problem (MTSP). Given a team of m agents (vehicles) that must visit n targets located in a real road network, a set of m optimized open paths (the Plan ) must be found such that each target is visited exactly once. The optimization objective is to minimize the driving distance of the path with the longest driving distance in the Plan (a Min-Max problem) so that this cooperative operation can be completed in the least amount of time. In several cases, the service area for each agent is limited (e.g., certain districts in a city). To simulate this real-world condition, two additional constraints are applied: the maximum geodetic range for each agent, and the maximum spatial range of targets for each agent. An easy-to-apply Genetic Algorithm (GA) with two novel initialization methods is presented to solve this route planning problem. In order to validate the developed route planning method and demonstrate its applicability, the method was tested in a real-world test case where it showed a decent performance. The results show that the applied limited service area constraints not only decrease the average driving distance of the longest route in the route plan, but also reduce the average runtime of the developed solution. Additionally, the performance of the proposed GA was benchmarked using 26 problems based on the TSPLIB instances, where the proposed GA achieved new Best Known Solution (BKS) values for 19 benchmark problems and a result equal to the BKS value for another two problems, demonstrating its superiority and robustness. The developed method can be used for route planning of the mobility and logistics services requiring multiple destinations to be reached by a fleet of vehicles, such as group ride-sharing and last-mile delivery.}, archive = {J_ASOC}, author = {Ali Maktabifard and Dávid Földes}, doi = {10.1016/j.asoc.2025.113959}, journal = {Applied Soft Computing}, month = {12}, pages = {113959}, shortjournal = {Appl. Soft. Comput.}, title = {Cooperative driving route planning based on asymmetric multi-agent path planning problem with limited service area constraints}, volume = {185}, year = {2025}, }
• (2025). A dual-branch self-supervised contrastive learning framework for emotion recognition based on time-frequency fusion. ASOC, 185, 113958. (www)

  @article{ , abstract = {Emotion recognition based on electroencephalography(EEG) signals is becoming a prominent research hotspot due to its wide-ranging applications in brain–computer interfaces (BCIs), mental health assessment, and human-computer interaction. Traditional emotion recognition methods often rely on supervised learning, which requires large amounts of labeled data to effectively train deep models. However, EEG signals exhibit inherent complexity and substantial variability across individuals and sessions, making it challenging to obtain consistent and reliable labels. In this paper, we propose a novel pretraining framework for EEG-based emotion recognition that enables mutual learning between time-domain and time-frequency-domain representations, while requiring simple network architectures. Experimental results demonstrate that our method achieves 84.39 % accuracy on the SEED dataset, and 89.01 % valence accuracy and 79.75 % arousal accuracy on the DEAP dataset using only 10 % labeled data, indicating strong performance under limited label conditions. Furthermore, we evaluate the transfer learning capability of our framework by pretraining it on the SEED dataset and then fine-tuning it on SEED-V. This cross-dataset transfer leads to a 1.9 % absolute improvement in classification accuracy on SEED-V, demonstrating the effectiveness of the learned representations in generalizing across datasets.}, archive = {J_ASOC}, author = {Jie Ouyang and Yangfan Liang and Hong Sun and Xianchao Zhang and Jingxue Chen and Gao Liu and Zhiquan Liu and Yining Liu}, doi = {10.1016/j.asoc.2025.113958}, journal = {Applied Soft Computing}, month = {12}, pages = {113958}, shortjournal = {Appl. Soft. Comput.}, title = {A dual-branch self-supervised contrastive learning framework for emotion recognition based on time-frequency fusion}, volume = {185}, year = {2025}, }
• (2025). A granularity time series forecasting model combining three-way decision and trend information granule. ASOC, 185, 113957. (www)

  @article{ , abstract = {Long-term forecasting of time series plays a vital role across diverse applications but is challenged by error accumulation arising from recursive predictions and the insufficient retention of trend information in conventional methods. To tackle these issues, we propose a novel forecasting model based on granular time series (GTS). The model utilizes an improved L 1 -trend filtering technique to achieve optimal segmentation of information granules, preserving essential trend features. Subsequently, we introduce dual evaluation functions based on distance similarity to jointly drive the three-way decision (TWD) process for aggregating information granules, thereby effectively reducing error propagation. Finally, the aggregated granules serve as inputs to a long short-term memory (LSTM) neural network to generate accurate forecasts. In addition, the proposed model is evaluated on several real-world datasets through sensitivity and comparative analyses. The results demonstrate that the model exhibits strong performance in long-term forecasting tasks.}, archive = {J_ASOC}, author = {Jianuan Qiu and Shuhua Su and Jingjing Qian}, doi = {10.1016/j.asoc.2025.113957}, journal = {Applied Soft Computing}, month = {12}, pages = {113957}, shortjournal = {Appl. Soft. Comput.}, title = {A granularity time series forecasting model combining three-way decision and trend information granule}, volume = {185}, year = {2025}, }
• (2025). Multi-objective genetic programming for binary classification with adaptive thresholds and a generalization-optimizing fitness function. ASOC, 185, 113956. (www)

  @article{ , abstract = {Genetic programming (GP) has been widely applied to classifier construction due to its flexible representation and powerful feature construction capabilities. Existing studies have proposed various fitness functions to improve GP-based classifiers, but most of them rely on a fixed decision threshold. However, when dealing with imbalanced classification problems, a fixed threshold often biases the model toward the majority class, thereby compromising overall performance. To address this issue, in this paper, we propose a novel multi-objective GP framework for constructing binary classifiers with adaptive threshold adjustment. During evolution, the method employs Youden’s Index to dynamically adjust the threshold of each individual, enabling the classifiers to better fit the underlying data distribution. In addition, we introduce a new class separation metric, dist t , to quantify the clarity of class boundaries and enhance the generalization ability of the evolved models. The framework jointly optimizes three objectives: minority class accuracy, majority class accuracy, and the proposed dist t metric. Experiments on 14 imbalanced datasets demonstrate that our method significantly outperforms conventional single-objective GP with fixed thresholds. Further results also confirm the positive impact of the proposed dist t metric on classification performance. Compared to seven existing GP algorithms and five traditional machine learning classifiers, our approach achieves superior overall performance and better generalization ability.}, archive = {J_ASOC}, author = {Minghui Bai and Yuan Gao and Xiaoying Gao and Jianbin Ma}, doi = {10.1016/j.asoc.2025.113956}, journal = {Applied Soft Computing}, month = {12}, pages = {113956}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-objective genetic programming for binary classification with adaptive thresholds and a generalization-optimizing fitness function}, volume = {185}, year = {2025}, }
• (2025). Adaptive fuzzy entropy optimization with opposition-based archimedes search for robust multilevel image segmentation. ASOC, 185, 113943. (www)

  @article{ , abstract = {Image segmentation plays a critical role in diverse computer vision applications. Multilevel thresholding (MLT) remains one of its most widely used unsupervised techniques due to its simplicity and interpretability. However, existing MLT methods often suffer from two major limitations: (1) the inability to adapt to local intensity variations and (2) the computational burden associated with high-dimensional threshold search. To address these challenges, this study proposes a novel segmentation framework that integrates a Proximity-Adaptive Fuzzy Entropy (PAFE) model with an Opposition-Based Learning-enhanced Archimedes Optimization Algorithm (OBL-EAOA). The PAFE model utilizes dynamically adjusted trapezoidal membership functions based on intensity proximity to candidate thresholds, allowing for a more adaptive and smooth entropy surface. Meanwhile, the OBL-EAOA enhances optimization performance through opposition-based learning and adaptive parameter control, improving exploration diversity and convergence speed. The proposed PAFE-EAOA framework is validated on two benchmark datasets, BSD500 and PASCAL VOC 2012, using five standard metrics: PSNR, SSIM, FSIM, SNR, and computation time. Compared with several state-of-the-art methods including Kapur Entropy (KE)-EAOA, Fuzzy Entropy (FE)-EAOA, Patch-Levy-Based Bees Algorithm (PLBA), Marine Predators Algorithm (MPA), Improved Grey Wolf Optimizer (IGWO), and standard Archimedes Optimization Algorithm (AOA), the proposed approach consistently achieves superior segmentation quality. Notably, it reduces computation time by up to 60 % and achieves statistically significant improvements, as confirmed by the Wilcoxon signed-rank test. These results demonstrate the framework’s robustness, scalability, and effectiveness for real-world MLT-based image segmentation.}, archive = {J_ASOC}, author = {Anusha Ganesan and Sungho Kim and Ganesan Nagabushnam}, doi = {10.1016/j.asoc.2025.113943}, journal = {Applied Soft Computing}, month = {12}, pages = {113943}, shortjournal = {Appl. Soft. Comput.}, title = {Adaptive fuzzy entropy optimization with opposition-based archimedes search for robust multilevel image segmentation}, volume = {185}, year = {2025}, }
• (2025). DT and LLM driven intelligent maintenance system for L-DED and DAG-based LLM fault diagnosis evaluation framework. ASOC, 185, 113942. (www)

  @article{ , abstract = {Metal additive manufacturing (AM) has revolutionized industries such as aerospace and automotive manufacturing due to its ability to rapidly prototype complex structures. Laser Directed Energy Deposition (L-DED) is a key AM technique, offering high deposition rates and superior mechanical properties. However, the inherent complexity and high cost of L-DED equipment demand reliable maintenance management to minimize downtime. Traditional maintenance approaches struggle to keep pace with escalating production demands and to cope with growing equipment complexity. To address this, we propose a dual-driven intelligent maintenance system for L-DED, integrating Digital Twins (DT) and Large Language Models (LLMs). The system features a comprehensive DT framework that synchronizes the virtual entity with the physical one in real time, it also incorporates an intelligent maintenance Q&A assistant powered by Retrieval-Augmented Generation (RAG), leveraging L-DED maintenance knowledge bases to provide accurate operational support. Additionally, we propose a Directed Acyclic Graphs (DAG)-based framework to assess LLMs’ ability to guide users through complete fault diagnosis. Our work aims to enhance the reliability and efficiency of L-DED maintenance through advanced digital technologies, ultimately improving productivity and reducing downtime in additive manufacturing.}, archive = {J_ASOC}, author = {Jian Tang and Shitong Peng and Jianan Guo and Danya Song and Dongna Gao and Weiwei Liu and Fengtao Wang}, doi = {10.1016/j.asoc.2025.113942}, journal = {Applied Soft Computing}, month = {12}, pages = {113942}, shortjournal = {Appl. Soft. Comput.}, title = {DT and LLM driven intelligent maintenance system for L-DED and DAG-based LLM fault diagnosis evaluation framework}, volume = {185}, year = {2025}, }
• (2025). A new large model with multi-scale feature fusion for fault diagnosis based on unified time series model. ASOC, 185, 113941. (www)

  @article{ , abstract = {Large model technology exemplified by large language models has been applied in the field of industrial fault diagnosis. However, existing large models are optimized for specific equipment types and have yet to fully exploit the potential of time-series monitoring data to enable widespread application across diverse mechanical equipment in various industrial scenarios. To address this challenge, a fault diagnosis large model (UniTS-FD) is designed based on unified time series model (UniTS). First, a multi-scale feature fusion backbone network is developed based on UniTS backbone to capture general mechanical fault features. Second, the fault classification head integrates the Pearson correlation coefficient to assess the similarity of class information within linear space for enabling adaptive classification. Third, P-LoRA fine-tuning approach incorporating LoRA and prompt technology is proposed to fine-tune the fault classification head, which enhances the generalization ability of the UniTS-FD model for fault diagnosis tasks of various mechanical equipment. Finally, the UniTS-FD model is pre-trained on 11 fault datasets and fine-tuning experiments were conducted on four different fault datasets to achieve cross-machine fault diagnosis. Experimental results demonstrate the effectiveness of the UniTS-FD in fault diagnosis tasks.}, archive = {J_ASOC}, author = {Zhiwei Zhang and Chengbin Wei and Weimin Zhang and Long Wen}, doi = {10.1016/j.asoc.2025.113941}, journal = {Applied Soft Computing}, month = {12}, pages = {113941}, shortjournal = {Appl. Soft. Comput.}, title = {A new large model with multi-scale feature fusion for fault diagnosis based on unified time series model}, volume = {185}, year = {2025}, }
• (2025). A federated learning-based method for personalized manufacturing service recommendation with collaborative relationships. ASOC, 185, 113940. (www)

  @article{ , abstract = {In the industrial Internet environment, the increasing complexity of manufacturing tasks has rendered them no longer accomplishable by independent manufacturing services. Meanwhile, current recommendation systems predominantly face challenges in maintaining data privacy and security during client parameter exchanges. To address these issues, this paper proposes CoFedSVD+ +, a federated learning-based method for personalized manufacturing service recommendation that integrates an enhanced SVD+ + algorithm with homomorphic encryption. First, we devise an enhanced similarity calculation method to analyze collaborative relationships among manufacturing services. Second, we implement a homomorphic encryption protocol within the federated learning framework to resolve data isolation challenges. Third, the improved SVD+ + algorithm is employed to capture implicit feedback information and predict missing Quality of Service (QoS) metrics. Fourth, a Top-N service composition recommendation list is generated through synergistic analysis of collaborative relationships and QoS predictions. Finally, we validate our approach using real-world case data from an industrial Internet platform. Experimental comparisons with existing recommendation algorithms demonstrate superior recommendation effectiveness of the proposed method.}, archive = {J_ASOC}, author = {Lei Wang and Jun Wang and Feng Xiang and Tongshun Li and Yang Xu and Yibing Li}, doi = {10.1016/j.asoc.2025.113940}, journal = {Applied Soft Computing}, month = {12}, pages = {113940}, shortjournal = {Appl. Soft. Comput.}, title = {A federated learning-based method for personalized manufacturing service recommendation with collaborative relationships}, volume = {185}, year = {2025}, }
• (2025). Span-level emotion-cause-category triplet extraction with instruction tuning LLMs and data augmentation. ASOC, 185, 113938. (www)

  @article{ , abstract = {Span-level emotion-cause-category triplet extraction is a fine-grained task in emotion cause analysis that aims to identify emotion spans, cause spans, and their corresponding emotion categories from documents. Existing methods, including clause-level emotion-cause pair extraction and span-level emotion-cause detection, often suffer from redundant information and difficulties in accurately classifying emotion categories, particularly when emotions are expressed implicitly or ambiguously. To overcome these challenges, this study explores a fine-grained approach to span-level emotion-cause-category triplet extraction and introduces an innovative framework that leverages instruction tuning and data augmentation techniques based on large language models. The proposed method employs task-specific triplet extraction instructions and utilizes low-rank adaptation to fine-tune large language models, eliminating the necessity for intricate task-specific architectures. Furthermore, an LLM-based data augmentation strategy is developed to address data scarcity by guiding large language models in generating high-quality synthetic training data. Extensive experimental evaluations demonstrate that the proposed approach significantly outperforms existing baseline methods, achieving at least a 12.8 % improvement in span-level emotion-cause-category triplet extraction metrics. The results demonstrate the method’s effectiveness and robustness, offering a promising avenue for advancing research in emotion cause analysis.}, archive = {J_ASOC}, author = {Xiangju Li and Dong Yang and Xiaogang Zhu and Faliang Huang and Peng Zhang and Zhongying Zhao}, doi = {10.1016/j.asoc.2025.113938}, journal = {Applied Soft Computing}, month = {12}, pages = {113938}, shortjournal = {Appl. Soft. Comput.}, title = {Span-level emotion-cause-category triplet extraction with instruction tuning LLMs and data augmentation}, volume = {185}, year = {2025}, }
• (2025). Deep cross-visual semantic hashing with self-calibrated collaborative attention. ASOC, 185, 113937. (www)

  @article{ , abstract = {Deep hashing has garnered considerable attention due to its remarkable retrieval efficiency and low storage cost, particularly in visual retrieval scenarios. However, current deep hashing methods generally integrate hash coding into a single-stream architecture, which limits the discriminative power of learned visual features and yields suboptimal hash codes. Additionally, over-reliance on semantic labels shared across samples fails to fully exploit the intrinsic semantic correlations between labels and corresponding visual features. To address these issues, we propose a deep cross-visual semantic hashing (DCvSH) method for image retrieval. First, we develop a visual image feature decoupling encoding network that leverages a self-calibrated collaborative attention mechanism to disentangle common and specific semantics across related images. These decoupled features are fed into a shared decoder for image reconstruction, yielding discriminative visual feature representations. Second, we construct a cross-visual semantic representation learning network with a two-level multi-layer perceptron to capture the underlying relationships between semantic label encodings and visual feature embeddings, while a hypergraph structure is introduced to preserve pairwise similarity relationships. Experimental results on the CIFAR-10, NUS-WIDE, and MIRFLICKR datasets demonstrate consistent improvements, with average mean average precision (mAP) scores reaching 0.895, 0.874, and 0.881 at different code lengths, respectively. Notably, DCvSH outperforms other baselines across all evaluation metrics.}, archive = {J_ASOC}, author = {Hao Feng and Xiangbo Zhou and Yue Wu and Jian Zhou and Banglei Zhao}, doi = {10.1016/j.asoc.2025.113937}, journal = {Applied Soft Computing}, month = {12}, pages = {113937}, shortjournal = {Appl. Soft. Comput.}, title = {Deep cross-visual semantic hashing with self-calibrated collaborative attention}, volume = {185}, year = {2025}, }
• (2025). Large model for fault diagnosis of industrial equipment based on a knowledge graph construction. ASOC, 185, 113936. (www)

  @article{ , abstract = {To address the significant heterogeneity of multi-modal data and the challenges in capturing fault semantics for industrial equipment, a fault diagnosis framework that integrates a time-frequency knowledge graph with the large model DeepSeek-V3 is proposed. Specifically, an unsupervised knowledge graph construction method is designed based on multi-modal vibration data signals. This method mines temporal evolution relationships using dynamic time warping and quantifies the relevance between features and faults via mutual information, thereby forming a dynamic graph representation. Additionally, DeepSeek-V3 encodes the natural language descriptions of vibration features, integrating graph structure and time-frequency map features to achieve collaborative reasoning and diagnosis among text, graphs, and maps. Experimental results show that the proposed method achieves high accuracy and significantly outperforms benchmark models, surpassing traditional methods. The proposed framework, through the deep integration of data-driven knowledge graphs and large model semantic understanding, demonstrates high precision, strong robustness, and transparent decision-making capabilities, providing new insights for intelligent diagnosis of industrial equipment.}, archive = {J_ASOC}, author = {Jichao Zhuang and Jiaming Yang and Weigang Li and Jian Chen and Yunjun Zheng and Zhuyun Chen}, doi = {10.1016/j.asoc.2025.113936}, journal = {Applied Soft Computing}, month = {12}, pages = {113936}, shortjournal = {Appl. Soft. Comput.}, title = {Large model for fault diagnosis of industrial equipment based on a knowledge graph construction}, volume = {185}, year = {2025}, }
• (2025). Interval-valued pythagorean fuzzy distance-based extended inferior ratio method for multiattribute decision-making: Application to green supplier selection in manufacturing industry. ASOC, 185, 113935. (www)

  @article{ , abstract = {Interval-valued Pythagorean fuzzy sets (IVPFSs) have emerged as a powerful tool for handling uncertainty and vagueness in multiattribute decision-making (MADM). In this paper, we first propose a novel distance measure for IVPFSs based on triangular divergence, which satisfies all core distance axioms and significantly improves discrimination ability compared to existing measures. Building on this, we introduce a maximizing deviation strategy with a new loss function to objectively determine attribute weights. Furthermore, we develop an extended inferior ratio (EIR) method that incorporates a dynamic weight parameter to flexibly balance the influence of positive and negative ideal solutions. The performance of the proposed method is demonstrated through a case study on green supplier selection in the manufacturing industry. The results indicate that, among the seven criteria evaluated, the most suitable suppliers are ranked as follows: β (1.0000), α (0.6471), δ (0.3500), ϵ (0.0690), and θ (0.0000). In addition, sensitivity and comparative analyses confirm the robustness and consistency of the proposed method, reflecting its effectiveness and practical value for sustainable decision-making in real-world scenarios.}, archive = {J_ASOC}, author = {Zhe Liu and Donglai Wang and Muhammet Deveci and Sukumar Letchmunan}, doi = {10.1016/j.asoc.2025.113935}, journal = {Applied Soft Computing}, month = {12}, pages = {113935}, shortjournal = {Appl. Soft. Comput.}, title = {Interval-valued pythagorean fuzzy distance-based extended inferior ratio method for multiattribute decision-making: Application to green supplier selection in manufacturing industry}, volume = {185}, year = {2025}, }
• (2025). Multidimensional trade performance assessment for integrating sustainability and economic diversification in OECD countries using a spherical fuzzy SIWEC-SPC-based decision support model. ASOC, 185, 113934. (www)

  @article{ , abstract = {Countries participate in regional trade blocks to enhance their trade activities and trade performance. This study proposes a novel approach to trade performance assessment, introducing innovations both in terms of performance parameters and methodology. From a performance parameter perspective, the inclusion of sustainable development and economic diversity levels in trade performance evaluation differentiates this study from traditional trade performance measurement approaches. From a methodological perspective, a hybrid model that simultaneously considers both subjective and objective criterion weighting based on expert opinions is introduced. The spherical fuzzy (SF)−simple weight calculation (SIWEC)−symmetry point of criterion (SPC)−opportunity losses-based polar coordinate distance (OPLO-POCOD) model, developed as a decision support system, is proposed for evaluating the trade performance of country blocs. SF-SIWEC is employed as the subjective criterion weighting method, while SPC serves as the objective criterion weighting method. The OPLO-POCOD alternative ranking method is employed to calculate the trade performance of countries. The proposed methodology is applied to the Organization for Economic Cooperation and Development (OECD) countries. According to the subjective weighting method, "sustainable development (0.0978)" emerges as the most important criterion, whereas in the objective weighting method, "number of product types (export) (0.1612)" is identified as the most significant one. In the final criterion weights, "number of product types (export) (0.1232)" is also determined to be the most important criterion. Considering the final criterion weights, the United States (0.9871) has the highest trade performance among the OECD countries. Thus, when both sustainability and economic diversification are considered, it is understood that the most influential criteria in determining multidimensional trade performance are the number of exported product types and sustainable development. From this perspective, the United States stands out as the country with the highest multidimensional trade performance among OECD countries. This study contributes to the literature by integrating sustainability and economic diversification parameters into trade performance measurement as well as proposing a comprehensive methodology for performance evaluation in trade blocs.}, archive = {J_ASOC}, author = {Galip Cihan Yalçın and Karahan Kara and Emre Kadir Özekenci and Vladimir Simic and Dragan Pamucar}, doi = {10.1016/j.asoc.2025.113934}, journal = {Applied Soft Computing}, month = {12}, pages = {113934}, shortjournal = {Appl. Soft. Comput.}, title = {Multidimensional trade performance assessment for integrating sustainability and economic diversification in OECD countries using a spherical fuzzy SIWEC-SPC-based decision support model}, volume = {185}, year = {2025}, }
• (2025). Optimizing supply routes in a network interdiction model with dual defense operations. ASOC, 185, 113933. (www)

  @article{ , abstract = {This study introduces a novel bi-level defender–attacker model (BDAM) designed to address real-world homeland defense scenarios. Building on the shortest path network interdiction problem (SPNIP), BDAM incorporates dual defense operations—node interdiction and edge destruction—while explicitly modeling the defender’s supply support. Unlike conventional SPNIP formulations, BDAM jointly considers the attacker’s path disruption and the defender’s logistical requirements, ensuring that all interdicted nodes are supported by available supply nodes without exceeding their capacity. To solve this NP-hard problem, a hybrid metaheuristic algorithm named Improved Simplified Swarm Optimization with Dijkstra (iSSOD) is proposed. The method integrates a population-based SSO framework with a randomized repair mechanism to ensure feasibility and an entropy-guided local search to enhance exploitation. The attacker’s optimal response is computed efficiently using Dijkstra’s algorithm, embedded within the defender’s fitness evaluation. The experimental results on 36 artificial instances demonstrate that iSSOD consistently outperforms several benchmark evolutionary algorithms, providing high-quality solutions through a defense-aware, supply-constrained optimization framework. Furthermore, a real-world case study based on geographic data validates the model’s applicability under realistic defense conditions.}, archive = {J_ASOC}, author = {Wei-Chang Yeh and Chyh-Ming Lai and Tsung-Hua Wu}, doi = {10.1016/j.asoc.2025.113933}, journal = {Applied Soft Computing}, month = {12}, pages = {113933}, shortjournal = {Appl. Soft. Comput.}, title = {Optimizing supply routes in a network interdiction model with dual defense operations}, volume = {185}, year = {2025}, }
• (2025). A continuous encoding-based representation for efficient multi-fidelity multi-objective neural architecture search. ASOC, 185, 113932. (www)

  @article{ , abstract = {Neural architecture search (NAS) is a powerful tool for automatically designing optimized deep learning models but is often limited by high computational costs, especially in multi-objective settings. To address this, an adaptive Co-Kriging-assisted multi-fidelity multi-objective NAS algorithm is proposed to reduce the computational cost and accelerate convergence of NAS by incorporating a clustering-based local multi-fidelity infill sampling strategy. Additionally, we introduce a novel continuous encoding method to compactly represent node connections within a generalized U-Net backbone, significantly reducing search dimensionality. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art NAS approaches under limited computational budgets on three benchmarks, a 2D Darcy flow regression task, a CHASE_DB1 biomedical image segmentation task, and an urban wind velocity prediction task. Analysis further shows that our algorithm autonomously identifies design patterns consistent with expert-curated U-Net variants in literature, confirming its efficiency and potential for insight into performant architectures.}, archive = {J_ASOC}, author = {Zhao Wei and Chin Chun Ooi and Yew-Soon Ong}, doi = {10.1016/j.asoc.2025.113932}, journal = {Applied Soft Computing}, month = {12}, pages = {113932}, shortjournal = {Appl. Soft. Comput.}, title = {A continuous encoding-based representation for efficient multi-fidelity multi-objective neural architecture search}, volume = {185}, year = {2025}, }
• (2025). Fall detection system using tabular GAN for data augmentation with integration of isolation forest model. ASOC, 185, 113931. (www)

  @article{ , abstract = {Effective Fall Detection Systems (FDS) are essential to minimize the risk of severe injuries among the elderly. However, the limited availability and imbalanced nature of fall detection data pose significant challenges to developing accurate models. This paper proposes a novel approach to enhance fall detection accuracy by using synthetic data generated through Tabular Generative Adversarial Network (GAN), combined with Isolation Forest and Autoencoder models. The dataset was augmented by factors of 5 and 10, and the models were evaluated using the area under the curve-receiver operating characteristic (AUC-ROC) and area under the curve–precision recall (AUC-PR) metrics. Notably, the Isolation Forest model improved from an AUC-ROC of 0.49 and AUC-PR of 0.43 (without augmentation) to 0.59 and 0.63, respectively, with 5x augmentation. Similarly, the Autoencoder showed an increase from 0.4 (AUC-ROC) and 0.2 (AUC-PR) to 0.5 and 0.54 with the same augmentation. These results demonstrate the effectiveness of synthetic data in improving anomaly detection performance. The findings suggest that advanced data augmentation techniques significantly improve FDS, thereby enhancing safety and quality of life for the vulnerable population.}, archive = {J_ASOC}, author = {Ali Nawaz and Najah Abu Ali and Amir Ahmad}, doi = {10.1016/j.asoc.2025.113931}, journal = {Applied Soft Computing}, month = {12}, pages = {113931}, shortjournal = {Appl. Soft. Comput.}, title = {Fall detection system using tabular GAN for data augmentation with integration of isolation forest model}, volume = {185}, year = {2025}, }
• (2025). Meta-learning-based adaptive operator selection for traveling salesman problem. ASOC, 185, 113930. (www)

  @article{ , abstract = {In evolutionary optimization, effectively leveraging knowledge about search operator performance is crucial for enhancing algorithmic results. Traditional operator selection strategies often rely on fixed heuristics or trial-and-error, which struggle to adapt to the nonstationary search dynamics of evolutionary runs—i.e., the stage-dependent, instance-dependent, and population-dependent shifts in operator effectiveness—and typically yield suboptimal performance. To address these challenges, we propose a novel meta-learning-based adaptive operator selection (AOS) framework. It leverages a Long Short-Term Memory (LSTM) neural network to learn temporal patterns of operator performance from historical data and dynamically adjust operator choice on-the-fly. The framework also integrates domain-specific biases to preserve population diversity and promote effective exploration, and it continuously updates its selection policy through dynamic online learning as the evolutionary process unfolds. Experiments on the Traveling Salesman Problem (TSP) benchmark demonstrate that the proposed LSTM-based AOS method significantly outperforms conventional approaches to operator selection. In particular, it achieved a median optimality gap of 9.87 % on a suite of TSP instances—approximately a 20 % improvement over the best fixed-operator configuration—indicating superior solution quality. Moreover, our approach consistently surpassed other state-of-the-art AOS techniques, underscoring the efficacy of the LSTM-driven framework and its significant potential to enhance evolutionary algorithm performance on complex optimization tasks.}, archive = {J_ASOC}, author = {Ho Young Jeong and Byung Duk Song}, doi = {10.1016/j.asoc.2025.113930}, journal = {Applied Soft Computing}, month = {12}, pages = {113930}, shortjournal = {Appl. Soft. Comput.}, title = {Meta-learning-based adaptive operator selection for traveling salesman problem}, volume = {185}, year = {2025}, }
• (2025). MACityChat: Integrating remote sensing professional large model with general-purpose large model for multi-domain urban land use analysis. ASOC, 185, 113929. (www)

  @article{ , abstract = {Urbanization remains a global trend, with urban land use being a key component of the process. The effective integration and management of land use are critical for the sustainable development of cities. Traditional urban land use analysis methods can fit dynamic models of land use changes nonlinearly, but they face two challenges: First, the analysis process of existing technologies is often a black-box, with unknown principles, reducing the reliability and authenticity of results. Second, traditional machine learning can only analyze urban land use changes from a single domain, such as remote sensing, overlooking the influence of economic and sociological factors. We propose an interpretable urban land use change analysis task and design MACityChat, a framework that combines remote sensing-specific large models with general-purpose large language models for multidisciplinary generalized analysis, while also visualizing the model’s analytical results. In this framework remote sensing images are input into a remote sensing large model, which transforms the semantic objects in the images into textual descriptions. These descriptions are then fed into a general-purpose large language model. A regional tag-guiding module directs the general-purpose language model to incorporate local economic, policy, and cultural knowledge to perform generalized analysis. Finally, the analysis results are visualized on the remote sensing images, providing a detailed examination of urban land use. Extensive experiments show that MACityChat can provide detailed and effective analyses of urban land use changes and visualize these analyses, offering an interpretable and superior solution to urban land use problems.}, archive = {J_ASOC}, author = {Tianyi Zhou and Zijie Huang and Hui Lin and Zhaobin Zhou and Jia Hu}, doi = {10.1016/j.asoc.2025.113929}, journal = {Applied Soft Computing}, month = {12}, pages = {113929}, shortjournal = {Appl. Soft. Comput.}, title = {MACityChat: Integrating remote sensing professional large model with general-purpose large model for multi-domain urban land use analysis}, volume = {185}, year = {2025}, }
• (2025). Causal inference and attribute correlation consistency for eliminating popularity bias in recommendation system. ASOC, 185, 113928. (www)

  @article{ , abstract = {Popularity bias has long been a persistent issue in recommendation systems, leading to misleading results and significant problems such as the Matthew Effect and the Information Cocoon Room. Existing studies have primarily focused on the elevation of long-tailed items, overlooking the crucial connection between users and items. Drawing inspiration from causal graphs, this paper introduces a novel framework called CIACC (Causal Inference and Attribute Correlation Consistency) to tackle the challenges posed by popularity bias. The framework leverages causal graphs to evaluate the compatibility between users and items and to gauge the influence of item popularity on rankings. It employs counterfactual inference to estimate the impact of item popularity on rankings and adheres to the consistent principle of attribute correlation to enhance the feature representation of long-tailed items. Through rigorous experiments conducted on three public datasets, we demonstrate that our CIACC framework outperforms state-of-the-art methods.}, archive = {J_ASOC}, author = {Bo Huang and Zhenhao Wang and Feifan Yan and Jin Liu and Hamido Fujita and Xiao Liu and Hanan Aljuaid}, doi = {10.1016/j.asoc.2025.113928}, journal = {Applied Soft Computing}, month = {12}, pages = {113928}, shortjournal = {Appl. Soft. Comput.}, title = {Causal inference and attribute correlation consistency for eliminating popularity bias in recommendation system}, volume = {185}, year = {2025}, }
• (2025). Strategy allocation for financial trading using competitive reinforcement learning and fuzzy logic. ASOC, 185, 113927. (www)

  @article{ , abstract = {Reinforcement learning has been widely adopted for quantitative trading and portfolio management. Research on reinforcement learning methodologies has simulated how a trading agent interacts with the financial market to realize goals, such as maximizing total profit or minimizing risk to optimize a portfolio within a specific period. Moreover, effective technical indicators have been developed for identifying price trends, but they often rely on a singular trading method, potentially limiting their effectiveness in dynamic, noisy markets. To address this limitation, we propose an intelligent, flexible trading system that employs multiple strategies, reflecting real-world trading practices. Our system, aimed at multistrategy allocation management, includes a global manager at a higher level and several local traders at a lower level. To address inherent noise in financial markets and prevent overfitting of the hyperparameters for each trading strategy, the system incorporates fuzzy logic that enhances its environmental representation. The system is tested in the Taiwan TXF and E-mini Standard and Poor’s 500 Index futures markets using 15-minute frequency data. The results demonstrate the system’s robustness and its effective market timing ability, marking a significant advancement in the field.}, archive = {J_ASOC}, author = {Ju-Chun Huang and Chiao-Ting Chen and Chih-Chung Chang and Szu-Hao Huang}, doi = {10.1016/j.asoc.2025.113927}, journal = {Applied Soft Computing}, month = {12}, pages = {113927}, shortjournal = {Appl. Soft. Comput.}, title = {Strategy allocation for financial trading using competitive reinforcement learning and fuzzy logic}, volume = {185}, year = {2025}, }
• (2025). A fast and robust ensemble evolving pixel cloud-based image segmentation approach. ASOC, 185, 113926. (www)

  @article{ , abstract = {The existing cluster-based image segmentation algorithms have the burden of iterative computation caused by the change of cluster centers and are sensitive to noise. In this paper, we present a fast and robust ensemble evolving pixel cloud-based image segmentation approach. The concept of pixel clouds by clustering pixels of the same pattern around their focal pixels is proposed. The following attributes distinguish the proposed algorithm: (1) The pixel clouds are evolvable according to the global densities of the incoming pixels and the number of pixel clouds is automatically determined. (2) The focal pixels of pixel clouds are dynamically updated with the highest local densities by using the recursive density estimation, which avoids redundant distance calculations when a new pixel arrives. (3) A multiscale morphological gradient reconstruction operation is employed to merge or filter meaningless pixel clouds, especially in noisy images, which helps to adaptively polish neighboring pixel clouds and compact the pixel clouds. (4) An ensemble structure is introduced to fasten the image segmentation speed by splitting the whole image into multiple independent sub-images, in which the pixel clouds are independently formed and evolved. Comprehensive experiments on natural images, remote sensing images and medical images reveal that the proposed approach surpasses the state-of-the-art algorithms in both segmentation accuracy and computational efficiency. Even for the noisy images, the proposed approach demonstrates more robust performance.}, archive = {J_ASOC}, author = {Tao Zhang and Hai-Jun Rong and Zhao-Xu Yang and Chi-Man Vong}, doi = {10.1016/j.asoc.2025.113926}, journal = {Applied Soft Computing}, month = {12}, pages = {113926}, shortjournal = {Appl. Soft. Comput.}, title = {A fast and robust ensemble evolving pixel cloud-based image segmentation approach}, volume = {185}, year = {2025}, }
• (2025). A data-driven approach to tackling academic stress-coping and mental health issues in college students using spherical fuzzy MARCOS methodology. ASOC, 185, 113925. (www)

  @article{ , abstract = {The drastically developing nature of the knowledge economy and the rising need for top-notch expertise have placed tremendous pressure on college students. As higher education becomes more accessible, masses of students are enrolling in colleges, which puts additional pressure on colleges and institutions; as a result, they cannot provide adequate resources to the students. As the class size increases, many students require mental health assistance, academic guidance, and financial aid, which then puts pressure on the teachers and the facilities. This flood of students overloads the facilities, resulting in it becoming more challenging to provide attention and concern, leading many students to feel overlooked and affecting their mental health. Due to not getting timely support, students may find it challenging to handle their academic responsibilities. Moreover, the students face a heavy workload, unclear guidance, and limited resource access. The objective of this study is to develop a structured, data-driven decision-making framework for systematically evaluating and improving student mental health and academic stress-coping strategies in a college setting. To address this, a comprehensive decision-making structure, measurement of alternatives, and ranking according to the compromised solution (MARCOS) within the spherical fuzzy (SF) environment, has been applied, which evaluates the key factors causing mental health issues by comparing the ideal and anti-ideal alternatives. The novelty of the proposed approach lies in leveraging the SF framework’s explicit ability to model hesitation (abstinence) alongside truth and falsity degrees, enabling more accurate representation of subjective psychological assessments compared to traditional fuzzy models. Furthermore, the method calculates utility functions corresponding to each alternative (coping technique), prioritizes the strategies, and selects the most effective intervention. The results reveal that personalized mental health plans emerged as the top-ranked coping strategy, highlighting the importance of tailored support in culturally and contextually diverse academic environments.}, archive = {J_ASOC}, author = {Raiha Imran and Munazza Amin and Kifayat Ullah and Dragan Pamucar and Zeeshan Ali and Oumaima Saidani and Vladimir Simic}, doi = {10.1016/j.asoc.2025.113925}, journal = {Applied Soft Computing}, month = {12}, pages = {113925}, shortjournal = {Appl. Soft. Comput.}, title = {A data-driven approach to tackling academic stress-coping and mental health issues in college students using spherical fuzzy MARCOS methodology}, volume = {185}, year = {2025}, }
• (2025). Secure and efficient vehicle control of autonomous vehicles using federated deep reinforcement learning. ASOC, 185, 113924. (www)

  @article{ , abstract = {Autonomous driving is largely considered a revolutionary technology and the ultimate alternative for smart urban mobility. Artificial Intelligence (AI) has been the main pillar of AV technology. However, despite the significant success achieved through the application of various Deep Reinforcement Learning (DRL) techniques in the field of AVs, numerous challenges remain to be addressed. Particularly, in the case of vehicle control, existing works have shown limited results when extensive action and state spaces are involved, and where the efficacy of agents is often compromised due to the variability of states and actions. In response to these challenges, we introduce a pioneering approach leveraging Federated Deep Reinforcement Learning (FDRL), which fosters the exchange of experiences among participating agents while safeguarding their privacy. FDRL enhances the exploration of all agents’ actions and states in different environments and effectively overcomes the problem of low sample efficiency in action and state spaces for all agents. Our approach is based on three distinct FDRL algorithms: Federated Proximal Policy Optimization (FPPO), Federated Deep Deterministic Policy Gradient (FDDPG), and Federated Deep Q-Network (FDQN), each tailored to control AVs within unique environments. Our results demonstrate the superiority of our approach in terms of average reward, waiting time, and speed, when compared to local models such as PPO, DDPG, and DQN.}, archive = {J_ASOC}, author = {Badr Ben Elallid and Nabil Benamar and Miloud Bagaa and Nabil Mrani}, doi = {10.1016/j.asoc.2025.113924}, journal = {Applied Soft Computing}, month = {12}, pages = {113924}, shortjournal = {Appl. Soft. Comput.}, title = {Secure and efficient vehicle control of autonomous vehicles using federated deep reinforcement learning}, volume = {185}, year = {2025}, }
• (2025). T5-based anomaly-behavior video captioning using semantic relation mining. ASOC, 185, 113923. (www)

  @article{ , abstract = {Video data consist of a series of images that change over time. The sequence of frames in a video provides important information on the motion and continuity of the video. Therefore, this dynamic information can be used to analyze the movement and behavior patterns of objects. Video captioning, which is used to explain a video, can describe the content of the video data and provide subtitles or descriptions in various languages. It can also explain the main points in a video with complex content, facilitating the information provided to users. In captioning, semantic analysis is used to identify the overall context of the data and generate the correct captions. However, captions are usually generated by focusing on major objects and actions, making it difficult to capture the details. In this paper, we propose text-to-text transfer transformer (T5)-based abnormal behavior video capturing using semantic relation mining. The proposed method generates captions with semantic features from video data based on environmental factors and improves the accuracy of video description by identifying the similarity of each caption for similar video and caption classification. This enables the classification and search of video data and is useful in video analysis systems, such as video monitoring and media analysis.}, archive = {J_ASOC}, author = {Min-Jeong Kim and Kyungyong Chung}, doi = {10.1016/j.asoc.2025.113923}, journal = {Applied Soft Computing}, month = {12}, pages = {113923}, shortjournal = {Appl. Soft. Comput.}, title = {T5-based anomaly-behavior video captioning using semantic relation mining}, volume = {185}, year = {2025}, }
• (2025). Simulating phenomenal consciousness using generative agents based on large language models. ASOC, 185, 113922. (www)

  @article{ , abstract = {Large Language Models (LLMs) still face challenges in tasks requiring understanding implicit instructions and applying common-sense knowledge. In such scenarios, LLMs may require multiple attempts to achieve human-level performance, potentially leading to inaccurate responses or inferences in practical environments, affecting their long-term consistency and behavior. This paper introduces the Internal Time-Consciousness Machine (ITCM), a computational consciousness structure to simulate the process of human consciousness. We further propose the ITCM-based Agent (ITCMA), which supports action generation and reasoning in open-world settings, and can independently complete tasks. ITCMA enhances LLMs’ ability to understand implicit instructions and apply common-sense knowledge by considering agents’ interaction and reasoning with the environment. The trained ITCMA performs better than state-of-the-art (SOTA) in the seen set. Even untrained ITCMA can achieve higher task completion rates than SOTA on the seen set, indicating its superiority over traditional intelligent agents in utility and generalization. In real-world tasks with quadruped robots, the task completion rate of untrained ITCMA is close to its performance in the unseen set, demonstrating its comparable utility and universality in real-world settings. CCS Concepts: ∙ Human-centered computing → Interactive systems and tools; ∙ Computing methodologies → Natural language processing.}, archive = {J_ASOC}, author = {Hanzhong Zhang and Jibin Yin and Haoyang Wang and Ziwei Xiang}, doi = {10.1016/j.asoc.2025.113922}, journal = {Applied Soft Computing}, month = {12}, pages = {113922}, shortjournal = {Appl. Soft. Comput.}, title = {Simulating phenomenal consciousness using generative agents based on large language models}, volume = {185}, year = {2025}, }
• (2025). Asymmetrical variable driven simulated binary crossover operator for large scale multi-objective optimization. ASOC, 185, 113921. (www)

  @article{ , abstract = {Due to the excessive number of decision variables, large scale multi-objective optimization has gradually become a hot research topic over the past few years. Although various reproduction operators have been proposed to effectively generate offspring solutions in the huge decision space, they have a common issue in that most of them are focused on the symmetrical variables and neglect asymmetrical variables. Based on the summary above, it is empirically proved that asymmetrical variables with great similarities play an important role in facilitating each other ′ s effective exploration of the huge decision space. Therefore, to effectively tackle large scale multi-objective optimization problems, the asymmetrical variables are incorporated into the design of a new asymmetrical variable driven simulated binary crossover operator. On the basis of the proposed asymmetrical variable driven simulated binary crossover operator, this paper further constructs a large scale multi-objective evolutionary framework based on asymmetrical variable driven simulated binary crossover operator. Extensive experiments and analyses on typical large scale multi-objective optimization problems with up to 10,000 decision variables and practical engineering problems with 6000 decision variables show its superiority over state-of-the-art optimizers.}, archive = {J_ASOC}, author = {Maoqing Zhang and Kun Wang}, doi = {10.1016/j.asoc.2025.113921}, journal = {Applied Soft Computing}, month = {12}, pages = {113921}, shortjournal = {Appl. Soft. Comput.}, title = {Asymmetrical variable driven simulated binary crossover operator for large scale multi-objective optimization}, volume = {185}, year = {2025}, }
• (2025). Towards agentic smart design: An industrial large model-driven human-in-the-loop agentic workflow for geometric modelling. ASOC, 185, 113920. (www)

  @article{ , abstract = {Agentic workflows, powered by Industrial Large Models (ILMs), represent a significant development emphasizing collaboration between humans and intelligent systems. This paper presents a structured perspective on the role of agentic workflows for smart design and manufacturing, grounded in integrating ILMs. We define an agentic workflow as a labeled Activity-on-Vertex (AOV) graph, where each node represents a functionally closed subtask and is executed by an ILM-based agent, a human operator, or an automated system. This formalism supports analyzable, modular, and hybrid execution, offering a foundation for modeling complex, mixed-initiative processes in manufacturing environments. To support real-world deployment, we introduce a set of reusable agentic workflow patterns that describe how ILM agents perceive, plan, and act in coordination with other components. Besides, a proof-of-concept case study illustrates the practical application of the human-in-the-loop framework through the agentic generation of CAD models. The study covers task decomposition, workflow implementation, and benchmarking, providing evidence for the feasibility of agentic workflows. Building upon these findings, this work contributes to advancing the development and application of agentic workflows in smart manufacturing contexts.}, archive = {J_ASOC}, author = {Keyou Zheng and Yuanwei Zhong and Xuyang Su and Jiewu Leng and Qiang Liu and Xin Chen}, doi = {10.1016/j.asoc.2025.113920}, journal = {Applied Soft Computing}, month = {12}, pages = {113920}, shortjournal = {Appl. Soft. Comput.}, title = {Towards agentic smart design: An industrial large model-driven human-in-the-loop agentic workflow for geometric modelling}, volume = {185}, year = {2025}, }
• (2025). Adaptive fault diagnosis of railway vehicle on-board controller with large language models. ASOC, 185, 113919. (www)

  @article{ , abstract = {Accurately identifying fault types of the Vehicle On-Board Controller (VOBC) in railway systems is of great significance for ensuring the safe operation of trains. Recently, Large Language Models (LLMs) have demonstrated excellent semantic understanding and natural language interaction capabilities, offering a novel solution for VOBC fault diagnosis. However, LLMs pretrained on general domains lack specific knowledge related to railway VOBC fault diagnosis scenarios, resulting in insufficient adaptability to railway-specific text corpora. This paper conducts an in-depth study on the adaptability of LLMs to VOBC fault diagnosis and proposes Railway Fault Diagnosis Large Language Model (RFD-LLM). First, we adopt railway domain adaptation based on Low-Rank Adaptation (LoRA) to match VOBC fault patterns. Second, instruction tuning is applied to achieve domain knowledge alignment and enhance the model’s ability to follow instructions. The proposed RFD-LLM is the first large language model-based fault diagnosis model for railway VOBC, capable of efficiently and accurately identifying seven types of VOBC fault patterns. RFD-LLM provides a new solution for the development of large models in the railway domain.}, archive = {J_ASOC}, author = {Cong Peng and Jiali Peng and Zisheng Wang and Zongyao Wang and Junjie Chen and Jianping Xuan and Tielin Shi}, doi = {10.1016/j.asoc.2025.113919}, journal = {Applied Soft Computing}, month = {12}, pages = {113919}, shortjournal = {Appl. Soft. Comput.}, title = {Adaptive fault diagnosis of railway vehicle on-board controller with large language models}, volume = {185}, year = {2025}, }
• (2025). An MADM model using frank operations based power aggregation operator under p,q-quasirung orthopair fuzzy sets for highway selection in war-plane landing. ASOC, 185, 113918. (www)

  @article{ , abstract = {In military logistics and operational planning, selecting an optimal highway for war-plane landings and take-offs is a critical and strategic decision. This process involves several key factors that directly affect mission success, operational safety, and public security. Among the most important attributes are the highway’s straight and long stretch with sufficient width to accommodate war-plane landing distances, and its surface condition, which must be free from obstacles, debris, and damage. Low traffic density is crucial to avoid the risk of collisions during landing. Additionally, favourable weather conditions, proximity to military camps, availability of emergency services and fuel, and a secure and hazard-free surrounding terrain are essential for safe and efficient operations. These factors collectively form the backbone of a reliable and tactical approach to highway selection for military air operations. Thus, in order to assess and rank various options for the landing and take-off of war planes, a strong and trustworthy procedure for making decisions is required. The purpose of this experiment is to build a comprehensive structure in multi-attribute decision making environment, using suggested p , q -quasirung orthopair fuzzy Frank power averaging as well as p , q -quasirung orthopair fuzzy Frank power geometric operators to capture ambiguity and uncertainty in highway selection. Furthermore, p , q -quasirung orthopair fuzzy Frank power weighted aggregation along with p , q -quasirung orthopair fuzzy Frank power weighted geometric operators are implemented for integrating the distance as well as similarity measures. Finally, sensitivity analysis and a comparison with the present technique are included to further demonstrate the superiority and validity of the technique that is suggested.}, archive = {J_ASOC}, author = {Sanjita Giri and Sankar Kumar Roy and Muhammet Deveci}, doi = {10.1016/j.asoc.2025.113918}, journal = {Applied Soft Computing}, month = {12}, pages = {113918}, shortjournal = {Appl. Soft. Comput.}, title = {An MADM model using frank operations based power aggregation operator under p,q-quasirung orthopair fuzzy sets for highway selection in war-plane landing}, volume = {185}, year = {2025}, }
• (2025). Real-time traffic flow optimization using large language models and reinforcement learning for smart urban mobility. ASOC, 185, 113917. (www)

  @article{ , abstract = {The exponential growth of metropolitan populations causes transportation network congestion, which increases fuel usage, travel time, and environmental damage. Traditional traffic management systems (TMS) seldom handle these issues in real time. Recently developed Large Language Models (LLMs), especially those using Reinforcement Learning (RL), may enhance urban transportation systems. Traffic management technology's real-time flexibility and shifting congestion patterns provide improved potential. Traditional approaches cannot estimate traffic flow or adapt to urban settings. A strong AI-driven method is needed to improve urban mobility and traffic flow. This paper introduces the LLM-RL Traffic Optimization Framework (LLM-RL-TOF). LLMs analyze real-time traffic data and give predictive insights in this context. Due to these new insights, the RL algorithm can improve traffic flow in real time and reduce congestion via dynamic traffic management. IoT sensors and urban traffic cameras capture real-time traffic data, including traffic volume and incidents. This data helps the LLM estimate bottlenecks, accidents, and traffic congestion. An RL agent uses LLM outputs to adjust traffic signal timing and suggest alternate routes. With real-time alternatives, traffic flow and urban mobility may be optimized. The junction throughput rate rose 17.5 %, the queue length accumulation index fell 22.3 %, and the average vehicle delay fell 18.6 %. The decrease in average vehicle delay enabled all these gains.}, archive = {J_ASOC}, author = {Arvind R. Singh and Muhammad Wasim Abbas Ashraf and Rajkumar Singh Rathore and Bin Li and M.S. Sujatha}, doi = {10.1016/j.asoc.2025.113917}, journal = {Applied Soft Computing}, month = {12}, pages = {113917}, shortjournal = {Appl. Soft. Comput.}, title = {Real-time traffic flow optimization using large language models and reinforcement learning for smart urban mobility}, volume = {185}, year = {2025}, }
• (2025). Progressive auto-encoding for domain adaptation in evolutionary multi-task optimization. ASOC, 185, 113916. (www)

  @article{ , abstract = {In recent years, evolutionary multi-task optimization (EMTO) has emerged as an effective paradigm for solving multiple optimization tasks simultaneously by leveraging knowledge transfer across tasks. The domain adaptation technique plays an important role in EMTO, as it helps align search spaces to support knowledge transfer among tasks. However, most existing domain adaptation methods rely on static pre-training or periodic re-matched mechanism, which do not adapt well to the dynamic change in evolving populations. In this paper, we propose a progressive auto-encoding (PAE) technique that enables continuous domain adaptation throughout the EMTO process. The PAE incorporates two complementary adaptation strategies: i) segmented PAE, which employs staged training of auto-encoders to achieve effective domain alignment across different optimization phases, and ii) smooth PAE, which utilizes eliminated solutions from the evolutionary process to facilitate more gradual and refined domain adaptation. We integrate the PAE into both single-objective and multi-objective multi-task evolutionary algorithms, yielding MTEA-PAE and MO-MTEA-PAE , respectively. Comprehensive experiments conducted on six benchmark suites and five real-world applications validate the effectiveness of our proposed PAE technique in enhancing domain adaptation capabilities within EMTO.}, archive = {J_ASOC}, author = {Qiong Gu and Yanchi Li and Wenyin Gong and Zhiyuan Yuan and Bin Ning and Chunyang Hu and Jicheng Wu}, doi = {10.1016/j.asoc.2025.113916}, journal = {Applied Soft Computing}, month = {12}, pages = {113916}, shortjournal = {Appl. Soft. Comput.}, title = {Progressive auto-encoding for domain adaptation in evolutionary multi-task optimization}, volume = {185}, year = {2025}, }
• (2025). Improving speech emotion recognition using gated cross-modal attention and multimodal homogeneous feature discrepancy learning. ASOC, 185, 113915. (www)

  @article{ , abstract = {Speech emotion recognition (SER) remains a significant and crucial challenge due to the complex and multifaceted nature of human emotions. To tackle this challenge, researchers strive to integrate information from diverse modalities through multimodal learning. However, existing multimodal fusion techniques often overlook the intricacies of interactions between different modalities, resulting in suboptimal feature representations. In this paper, we propose WavFusion, a multimodal framework designed for SER that tackles key research challenges, such as effective multimodal fusion, modality heterogeneity, and discriminative representation learning. By utilizing a gated cross-modal attention mechanism and multimodal homogeneous feature discrepancy learning, WavFusion outperforms existing state-of-the-art methods on benchmark datasets. Our research highlights the importance of capturing subtle cross-modal interactions and learning discriminative representations for accurate multimodal SER. Experimental results indicate that the proposed method is highly competitive and better than most of the latest state-of-the-art methods for SER. WavFusion achieves 0.78 % and 1.27 % improvement in accuracy and 0.74 % and 0.44 % improvement in weighted F1 score over the previous methods on the IEMOCAP and MELD datasets, respectively.}, archive = {J_ASOC}, author = {Feng Li and Jiusong Luo and Wanjun Xia}, doi = {10.1016/j.asoc.2025.113915}, journal = {Applied Soft Computing}, month = {12}, pages = {113915}, shortjournal = {Appl. Soft. Comput.}, title = {Improving speech emotion recognition using gated cross-modal attention and multimodal homogeneous feature discrepancy learning}, volume = {185}, year = {2025}, }
• (2025). Wind turbine blades defect detection based on global and local attention with multi-feature fusion. ASOC, 185, 113914. (www)

  @article{ , abstract = {Wind turbine blades are prone to small-scale defects—such as cracks, corrosion, and contamination—during long-term operation. Accurate detection of these defects is essential for ensuring the safety and efficiency of wind power systems. However, small-object detection remains challenging due to limited feature representation and weak discriminative cues. To address this, an enhanced YOLOX-s-based framework called Global-Frequency Dual-aware YOLOX (GFD-YOLOX) is proposed. GFD-YOLOX introduces three main improvements. First, the Path Aggregation Feature Pyramid Network (PAFPN) in the neck is replaced with Dual-Frequency Fused Bidirectional Feature Pyramid Network (DFF-BiFPN) to strengthen multi-scale contextual representation. Second, the backbone bottleneck is redesigned with a lightweight structure, improving computational efficiency and convergence speed. Third, a Hierarchical Frequency-Adaptive Fusion (HFAF) module is integrated to enhance cross-scale feature interaction by combining fine-grained and global information. On the self-constructed WTBlade-Defect dataset (3570 annotated images, five defect types: corrosion, hide-craze, surface-eye, thunderstrike, dirt), GFD-YOLOX achieves mAP@0.5 and mAP@0.5:0.95 scores of 94.5 % and 68.9 %, respectively, with 44.3 FPS inference—improving by 13.6 % and 14.4 % over state-of-the-art models. On the public dataset of Ashley Foster et al., it achieves 94.8 % and 69.3 %, with gains of 10.4 % and 10.9 %. These results demonstrate that GFD-YOLOX delivers substantial accuracy gains while maintaining real-time speed and strong cross-dataset generalization, indicating high potential for deployment in operational wind turbine inspection systems.}, archive = {J_ASOC}, author = {Dandan Liu and Mingjie Liu}, doi = {10.1016/j.asoc.2025.113914}, journal = {Applied Soft Computing}, month = {12}, pages = {113914}, shortjournal = {Appl. Soft. Comput.}, title = {Wind turbine blades defect detection based on global and local attention with multi-feature fusion}, volume = {185}, year = {2025}, }
• (2025). Heterogeneous feature selection with group structure mining in fuzzy decision systems for medical diagnosis. ASOC, 185, 113913. (www)

  @article{ , abstract = {In practical applications such as medical diagnosis and group decision making, the potential structural information contained in multi-dimensional features in the form of group domains plays an important role. However, most existing feature selection methods adopt transformed feature spaces for group structure analysis, which lack intrinsic semantic information interpretation. Meanwhile, fuzzy and uncertain heterogeneous data acquired from multiple devices increase the difficulty of task learning. Motivated by these two issues, this work devises a Heterogeneous Feature Selection method with Group Structure Mining in fuzzy decision systems (HFS-GSM), which follows the principle of one “strategy" and one “mechanism". Specifically, a feature group generation strategy based on fuzzy approximation Markov blanket is first designed for mining features with group structure, which introduces the concept of Markov blanket into the fuzzy rough set and utilizes the idea of approximation and fuzzy uncertainty measures. Then, a fuzzy dependency-based overlapping group elimination mechanism is proposed by attribution division, which avoids local redundancy while preserving global discriminative information. Furthermore, the effectiveness of HFS-GSM is verified in comparison with seven representative feature selection methods on publicly available medical datasets. Finally, medical diagnosis data provided by a hospital are obtained to demonstrate the reliability and utility of HFS-GSM in practical applications.}, archive = {J_ASOC}, author = {Jihong Wan and Hongmei Chen and Li Xiao and Chuangpeng Shen and Wei Huang and Xiaoping Li}, doi = {10.1016/j.asoc.2025.113913}, journal = {Applied Soft Computing}, month = {12}, pages = {113913}, shortjournal = {Appl. Soft. Comput.}, title = {Heterogeneous feature selection with group structure mining in fuzzy decision systems for medical diagnosis}, volume = {185}, year = {2025}, }
• (2025). Ahybrid ICEEMDAN-BO-GRU model for real-time grouting parameter prediction. ASOC, 185, 113912. (www)

  @article{ , abstract = {Unit injection volume is one of the core control indicators for evaluating the construction quality of curtain grouting. In actual construction processes, accurate prediction of unit injection volume enables early warning of abnormal conditions during construction. However, geological conditions and construction parameters significantly influence unit injection volume, among other factors, and the data exhibits high nonlinearity. Previous studies have achieved suboptimal prediction accuracy. To address the issue of low prediction accuracy in traditional algorithms when handling highly nonlinear data, this study first applies the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) algorithm to decompose the raw data, and the nonlinear features in the curtain grouting data are effectively eliminated. Subsequently, the Gated Recurrent Unit (GRU) and Bayesian Optimization (BO) algorithms are organically integrated, and Bayesian optimization is applied to the hyperparameters of the GRU model to enhance its predictive performance, thereby constructing an intelligent predictive model for the curtain grouting unit injection volume. Finally, an engineering application study was conducted using an actual curtain grouting project at a pumped-storage power station. The results indicate that the model exhibits good computational accuracy and efficiency, providing theoretical and methodological support for constructing intelligent prediction models in curtain grouting.}, archive = {J_ASOC}, author = {Fei Tong and Dongqing Bai and Xufei Ma and Lin Cheng and Jie Yang and Xiangyu Cao}, doi = {10.1016/j.asoc.2025.113912}, journal = {Applied Soft Computing}, month = {12}, pages = {113912}, shortjournal = {Appl. Soft. Comput.}, title = {Ahybrid ICEEMDAN-BO-GRU model for real-time grouting parameter prediction}, volume = {185}, year = {2025}, }
• (2025). DefGCL: Defence-enhanced graph contrastive learning against attribute inference attacks. ASOC, 185, 113911. (www)

  @article{ , abstract = {Graph-structured data are prevalent in many real-world applications, such as social networks, drug discovery, and fraud detection. While Graph Neural Networks (GNNs) have shown remarkable performance by capturing rich relational patterns, their success often relies on large labeled datasets and raises growing privacy concerns. Graph Contrastive Learning (GCL) has emerged as a powerful unsupervised alternative by leveraging data augmentations to learn robust representations without labeled data. However, recent studies reveal that GCL models are particularly vulnerable to attribute inference attacks, and existing works prioritize performance improvement over privacy protection. To address this issue, we propose a D e f ense-enhanced G raph C ontrastive L earning, dubbed DefGCL , that integrates four coordinated defense strategies to enhance privacy without degrading utility. Specifically, DefGCL employs edge-based graph augmentations to limit exposure to structural attributes, selects negative samples with low attribute sensitivity scores to reduce leakage, modifies the contrastive loss to decouple graph embeddings from attributes, and injects differential privacy noise during the embedding stage. Extensive experiments on five benchmark datasets demonstrate that DefGCL achieves state-of-the-art (SOTA) performance in both privacy preservation and task accuracy. For instance, on the AIDS dataset, DefGCL reduces attribute inference accuracy by 35 % while incurring only a 0.60 % drop in main task performance. Additionally, DefGCL improves computational efficiency by reducing runtime by nearly 50 % compared to baseline methods. 1}, archive = {J_ASOC}, author = {Jinyin Chen and Fanyu Ao and Wenbo Mu and Haiyang Xiong}, doi = {10.1016/j.asoc.2025.113911}, journal = {Applied Soft Computing}, month = {12}, pages = {113911}, shortjournal = {Appl. Soft. Comput.}, title = {DefGCL: Defence-enhanced graph contrastive learning against attribute inference attacks}, volume = {185}, year = {2025}, }
• (2025). Interval multi-feature sparse transformer-CNN: A synergistic approach to precise and efficient spatio-temporal wind speed interval-value prediction. ASOC, 185, 113910. (www)

  @article{ , abstract = {Accurate and reliable wind speed prediction is critical for stabilizing wind power integration. However, current methods are limited by accuracy and stability issues, hindering their large-scale application in wind farms. To overcome these problems, this study innovatively proposes an IMFSformer-CNN model integrating three core components. First, the spatio-temporal and multi-factor feature extraction technology comprehensively captures the spatio-temporal patterns and complex dependencies of wind speed dynamics, incorporating multiple factors such as meteorological variables and spatial correlation. Second, the multi-feature sparse attention mechanism reduces computational complexity by combining sparse attention with multi-feature attention, enhancing representation ability and scalability for precise interval value prediction. Finally, the enhanced interval spatio-temporal prediction fusion model combines the global dependency modeling capabilities of the improved Transformer architecture with the local receptive field advantages of CNN. This hybrid design facilitates the simultaneous capture of both macro-scale atmospheric patterns and micro-scale wind speed fluctuations. The model achieved prediction interval coverage probabilities of 0.921 and 0.899, and coverage width criteria of 1.493 and 3.776, outperforming other models on both datasets. This significantly enhances accuracy and practical value for wind farm cluster forecasting, supporting more reliable and efficient wind energy integration into power grids.}, archive = {J_ASOC}, author = {Weiyi Jiang and Jujie Wang and Xuecheng He}, doi = {10.1016/j.asoc.2025.113910}, journal = {Applied Soft Computing}, month = {12}, pages = {113910}, shortjournal = {Appl. Soft. Comput.}, title = {Interval multi-feature sparse transformer-CNN: A synergistic approach to precise and efficient spatio-temporal wind speed interval-value prediction}, volume = {185}, year = {2025}, }
• (2025). Multi-granularity spatiotemporal fusion neural network with denoising reconstruction for human motion prediction. ASOC, 185, 113909. (www)

  @article{ , abstract = {Human motion prediction has significant application value in many fields. However, existing methods often fail to fully capture the spatial relationships between joints and the temporal flow of information when modeling complex spatiotemporal dependencies. Additionally, these methods are prone to overfitting dominant features while neglecting other important aspects, and struggle with perceiving contour features effectively. To address these issues, this study introduces a novel encoder-decoder framework. The encoder generates a dual-layer adaptive adjacency matrix using a distance partition strategy to parameterize joint relationships, while incorporating a gating mechanism to control the temporal flow of information. The decoder then employs separate spatiotemporal attention modules to decode temporal and spatial features independently. These features are subsequently reconstructed through a spatiotemporal fusion strategy, effectively decoupling and modeling complex spatiotemporal dependencies. To address the issue of overfitting to dominant features, we introduce a denoising reconstruction strategy that allows the model to learn richer combinations of spatiotemporal features under multiple constraints. Furthermore, a multi-granularity information adaptive fusion module is incorporated to achieve adaptive fusion of both local and contour features. Experimental results across several benchmark datasets demonstrate that our method significantly outperforms the state-of-the-art approaches, showcasing its effectiveness in human motion prediction tasks.}, archive = {J_ASOC}, author = {Yong Li and Linfeng Zhu and Haofei Xie and Xinchang Yi}, doi = {10.1016/j.asoc.2025.113909}, journal = {Applied Soft Computing}, month = {12}, pages = {113909}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-granularity spatiotemporal fusion neural network with denoising reconstruction for human motion prediction}, volume = {185}, year = {2025}, }
• (2025). FFNN: Fractional order basis function multi-step neural network method for fractional partial differential equations. ASOC, 185, 113907. (www)

  @article{ , abstract = {With the advancement in artificial intelligence technology, the increasing number of researchers utilize it to address complex equations in ocean engineering. So the technology of artificial intelligence has become a practical area of research. In this paper, we design a novel method to solve the fractional order long water wave equation, which is called the fractional order basis function multi-step neural network. Firstly, a power series is constructed based on a fractional order basis function, which serves as the approximate solution. Secondly, neural networks and the initial conditions of differential equations are integrated into the construction of approximate solutions. Furthermore, the solution is discretized, and a multi-step unfolding strategy is employed on the resulting discrete solution. This approach ensures that each point in the solution is influenced by its predecessor. By means of repeated applications of the optimization algorithm, the residuals are successively diminished, thereby yielding approximate solutions to the equations. Finally, the efficacy and versatility of the proposed strategy were validated through a series of numerical experiments. Compared with the method of fractional physics-informed neural networks, there are up to 18.7 -fold and 22.8 -fold increases in stability of average and maximum residuals. Simultaneously, initial conditions are retained in new solutions.}, archive = {J_ASOC}, author = {Jianke Zhang and Xudong Tian and Chang Zhou}, doi = {10.1016/j.asoc.2025.113907}, journal = {Applied Soft Computing}, month = {12}, pages = {113907}, shortjournal = {Appl. Soft. Comput.}, title = {FFNN: Fractional order basis function multi-step neural network method for fractional partial differential equations}, volume = {185}, year = {2025}, }
• (2025). MCSMOTE: A transition matrix-driven oversampling technique for class imbalance. ASOC, 185, 113906. (www)

  @article{ , abstract = {Class imbalance presents a challenge in machine learning, often skewing predictive performance toward the majority class and undermining the accuracy of minority class predictions. To address this, we introduce MCSMOTE, a novel resampling method that employs a transition matrix-based Monte Carlo mechanism for generating synthetic samples. MCSMOTE differentiates itself by modeling the relationships among features and leveraging probabilistic transitions to generate synthetic data points that effectively capture the underlying data structure. This approach ensures enhanced representativeness of the minority class while approximating the local structure of the minority class and thereby generating samples that reflect the underlying data patterns. Comprehensive experiments across 63 diverse imbalanced datasets demonstrate that MCSMOTE consistently outperforms nine widely used resampling techniques—NORES, ROS, SMOTE, ADASYN, BLSMOTE, RWO, SMOTEWB, DeepSMOTE, RWO, and GQEO—when evaluated using multiple classifiers and six key performance metrics: balanced accuracy, F1-score, G-mean, MCC, ROCAUC, and ROCAUC. Results show that MCSMOTE achieves the highest average performance across all metrics. Friedman and Nemenyi tests confirm that these improvements are statistically significant. An ablation study further highlights the stability and effectiveness of MCSMOTE’s hyperparameter choices across different data characteristics. These findings establish MCSMOTE as a powerful and reliable solution for addressing class imbalance in machine learning applications.}, archive = {J_ASOC}, author = {Fatih Sağlam and Mehmet Ali Cengiz}, doi = {10.1016/j.asoc.2025.113906}, journal = {Applied Soft Computing}, month = {12}, pages = {113906}, shortjournal = {Appl. Soft. Comput.}, title = {MCSMOTE: A transition matrix-driven oversampling technique for class imbalance}, volume = {185}, year = {2025}, }
• (2025). Limited-budget consensus with maximum consensus level for group decision making. ASOC, 185, 113905. (www)

  @article{ , abstract = {Group decision making usually requires in-depth discussions to form a consensus acceptable to the entire group, which has attracted more and more research in recent years. Despite extensive studies on soft consensus, the relationship between consensus level and consensus cost remains unclear. This study establishes—for the first time—a precise mathematical relationship demonstrating that higher consensus levels require proportionally greater consensus costs. This finding provides critical theoretical grounding for consensus modeling. Recognizing that the cost of achieving consensus cannot be infinite and must be within a certain budget, we develop a model to determine the maximum achievable consensus level under a limited-budget. The consensus model with non-cooperators is also explored and formulated. The proposed models are applied to online lending platforms, providing a practical framework for measuring consensus levels and achieving soft consensus between lenders and borrowers within a limited-budget. This work contributes to understanding the relationship between consensus level and consensus cost, as well as the achievement of the maximum possible consensus level under limited-budget, which is relevant in scenarios where resources are finite.}, archive = {J_ASOC}, author = {Huanhuan Zhang and Dongjie Guo and Yifeng Ma}, doi = {10.1016/j.asoc.2025.113905}, journal = {Applied Soft Computing}, month = {12}, pages = {113905}, shortjournal = {Appl. Soft. Comput.}, title = {Limited-budget consensus with maximum consensus level for group decision making}, volume = {185}, year = {2025}, }
• (2025). Discrimination index based attribute reduction for partially labeled heterogeneous data via a new prediction label strategy and statistical distribution of data. ASOC, 185, 113904. (www)

  @article{ , abstract = {Existing methods often struggle to effectively utilize the statistical distribution of data and unlabeled information when dealing with partially labeled heterogeneous data, leading to limited performance in attribute reduction. To address this issue, this paper presents a novel semi-supervised attribute reduction method that integrates statistical distribution of data, conditional discrimination index and a new prediction label strategy. Initially, the similarity between objects is constructed by analyzing the statistical distribution of data. Subsequently, a new prediction label strategy is introduced, which not only utilizes existing label information but also completes missing information by predicting new labels, thereby enhancing the completeness and usability of data. Finally, conditional discrimination index, a tool for measuring the importance of attributes in classification tasks, is employed to design an attribute reduction algorithm. The experimental results on real-world partially labeled heterogeneous datasets show that the proposed p-CDIAR algorithm exhibits superior performance compared to the existing nine reduction algorithms, with an average increase of 4.67 % in classification accuracy and 6.32 % in outlier detection performance.}, archive = {J_ASOC}, author = {Qingnian Li and Haixin Huang and Tao Lu and Huaming Wei and Zhaowen Li}, doi = {10.1016/j.asoc.2025.113904}, journal = {Applied Soft Computing}, month = {12}, pages = {113904}, shortjournal = {Appl. Soft. Comput.}, title = {Discrimination index based attribute reduction for partially labeled heterogeneous data via a new prediction label strategy and statistical distribution of data}, volume = {185}, year = {2025}, }
• (2025). An unsupervised framework for drift-aware anomaly detection in streaming time series. ASOC, 185, 113903. (www)

  @article{ , abstract = {This paper presents an unsupervised adaptive drift-aware anomaly detection framework (ADA-ADF) designed to address the challenges of concept drift in time series data streams. ADA-ADF integrates a hybrid drift detection mechanism, combining statistical tests with performance-based metrics to accurately identify and distinguish between sudden and incremental drifts. To ensure effective adaptation, it employs a replay-based model update strategy that adjusts replay ratios in a drift-specific manner and incorporates representative historical data based on reconstruction errors. This approach allows the model to seamlessly adapt to evolving data distributions while maintaining high stability and accuracy. Extensive experiments on four diverse datasets demonstrate ADA-ADF’s superior performance in managing various drift and application scenarios. It consistently outperforms state-of-the-art methods, particularly in environments characterized by incremental or sudden drifts. With robust adaptability to changing data patterns and accurate anomaly detection capabilities, ADA-ADF provides a reliable solution for real-world applications, such as IoT and environmental monitoring.}, archive = {J_ASOC}, author = {Danlei Li and Nirmal-Kumar C. Nair and Kevin I-Kai Wang}, doi = {10.1016/j.asoc.2025.113903}, journal = {Applied Soft Computing}, month = {12}, pages = {113903}, shortjournal = {Appl. Soft. Comput.}, title = {An unsupervised framework for drift-aware anomaly detection in streaming time series}, volume = {185}, year = {2025}, }
• (2025). A MADM framework for classifying wireless sensor networks in industrial automation and monitoring using hesitant bipolar complex fuzzy dombi power operators. ASOC, 185, 113902. (www)

  @article{ , abstract = {Wireless sensor networks (WSNs) are essential for industrial automation and monitoring because they eliminate the need for wires by enabling real-time data collection, processing, and transmission. They are highly effective for a wide range of applications, from optimizing industrial processes to monitoring machinery and environmental conditions. Using standard methods for evaluating WSNs is insufficient to identify many of the features that are important for making decisions. However, a major flaw in the existing approaches is their inability to simultaneously consider the negative features and extra information associated with attributes. There may be a greater chance of significant information loss as a result. Keeping in mind all those flaws, this manuscript has formulated a new methodology for the evaluation of the best WSNs classification. The hesitant bipolar complex fuzzy (HBCF) Dombi power approach is probably different from all other existing approaches because of its different affordable characteristics of hesitancy nature with dual aspects and saving extra information without any loss. Moreover, this novel approach can easily manage the bipolar nature of any information. Also, in this manuscript, a case study focused on WSNs has been examined. The comparative analysis and results are also discussed, which demonstrate the superiority and effectiveness of the proposed work.}, archive = {J_ASOC}, author = {Tahir Mahmood and Hafiz Muhammad Waqas and Ubaid ur Rehman}, doi = {10.1016/j.asoc.2025.113902}, journal = {Applied Soft Computing}, month = {12}, pages = {113902}, shortjournal = {Appl. Soft. Comput.}, title = {A MADM framework for classifying wireless sensor networks in industrial automation and monitoring using hesitant bipolar complex fuzzy dombi power operators}, volume = {185}, year = {2025}, }
• (2025). Extended belief rule-based system with online joint learning strategy. ASOC, 185, 113901. (www)

  @article{ , abstract = {With the proliferation of dynamic data streams in the advanced technology environment, it is necessary to solve the evolving classification problems by using adaptable and interpretable artificial intelligence techniques. To meet this challenge, a new extended belief rule-based (EBRB) system incorporating online joint learning strategy is proposed in this paper. The online joint learning strategy comprises two key components: rule update and parameter update schemes. In the rule update scheme, different rule incorporation processes are designed for the labeled or unlabeled input data while overlapping and redundant rules are removed from the rule base. To adapt the updated rule base, the parameter update scheme is designed to retune the parameters within updated rule base. The antecedent attribute weights are optimized using the Bayesian optimization algorithm and the rule weights are updated based on the consistency of rules. To evaluate the performance of the developed system, it is applied to assist radiologists in diagnosing thyroid nodules. Compared with the existing offline EBRB systems and online learning methods, the proposed online joint learning EBRB system could generate higher classification accuracy with fewer rules in the limited running time.}, archive = {J_ASOC}, author = {Bingbing Hou and Min Xue and Leilei Chang and Zijian Wu}, doi = {10.1016/j.asoc.2025.113901}, journal = {Applied Soft Computing}, month = {12}, pages = {113901}, shortjournal = {Appl. Soft. Comput.}, title = {Extended belief rule-based system with online joint learning strategy}, volume = {185}, year = {2025}, }
• (2025). A collaborative competition multitasking framework for constrained multi-objective optimization. ASOC, 185, 113900. (www)

  @article{ , abstract = {Constrained multi-objective optimization problems (CMOPs) are common in the real world. Constrained multi-objective evolutionary algorithms (CMOEAs) based on evolutionary multi-tasking show excellent performance in solving CMOPs. However, not all tasks can find useful information during the process of evolution, which inevitably results in a waste of computing resources. In this paper, a CMOEA based on collaborative competition multitasking (TCCMT) is proposed, in which two auxiliary tasks are constructed to co-evolve with the main task in a collaborative competition manner. During the process of evolution, only the dominant auxiliary task is selected to help the main task evolve, which reduces the resource consumption to evolve the invalid tasks. Meanwhile, the evolutionary process is divided into three stages in order to balance exploration and exploitation. The auxiliary tasks customize the constrained adaptive regression strategy and double angle enhancement strategy respectively to improve the ability to solve different problems. Compared with the nine most advanced CMOEAs on 33 benchmark problems and 7 real-world engineering problems, the Friedman test results show that TCCMT achieves the best rank on all test problems and exhibits a statistically significant difference.}, archive = {J_ASOC}, author = {Xinyu Feng and Qianlong Dang and Xiaochuan Gao and Yanghui Wu and Lifei Zheng}, doi = {10.1016/j.asoc.2025.113900}, journal = {Applied Soft Computing}, month = {12}, pages = {113900}, shortjournal = {Appl. Soft. Comput.}, title = {A collaborative competition multitasking framework for constrained multi-objective optimization}, volume = {185}, year = {2025}, }
• (2025). A Q-learning based genetic algorithm for collaborative optimization of import container allocation and yard crane deployment in container terminal. ASOC, 185, 113899. (www)

  @article{ , abstract = {The implementation of truck appointment systems in container terminals has led to a greater focus on optimizing terminal operations using appointment information. Container allocation and yard crane deployment problems are the primary issues of yard management that are influenced by truck appointment information. This paper investigates the impact of appointment information on yard operation management, specifically focusing on the container allocation for import tasks and optimizing the deployment of yard cranes based on the workload. A collaborative optimization model is proposed to minimize the workload imbalance across blocks. This model considers the impact of truck appointment information, allocating storage space and storage numbers for import tasks, as well as the minimum number of transfers to arrange yard cranes. Due to the complexity of the model, a self-learning genetic algorithm utilizing Q-learning is designed to solve the proposed model. A comparison was made between the optimization model that takes truck appointments into account and the traditional one which did not. The results indicate that the workload imbalance can be decreased by a minimum of 22.06 % when considering information on the truck appointments. The experimental results demonstrate that the proposed approach successfully reduces workload imbalance among blocks and mitigates the negative effects caused by yard congestion.}, archive = {J_ASOC}, author = {CuiJie Diao and Ying Huang and KangZhen Peng and ZhiHong Jin}, doi = {10.1016/j.asoc.2025.113899}, journal = {Applied Soft Computing}, month = {12}, pages = {113899}, shortjournal = {Appl. Soft. Comput.}, title = {A Q-learning based genetic algorithm for collaborative optimization of import container allocation and yard crane deployment in container terminal}, volume = {185}, year = {2025}, }
• (2025). Fusing local density and approximate distance for nonparametric outlier detection. ASOC, 185, 113898. (www)

  @article{ , abstract = {Outlier detection is an essential yet challenging task in intelligent data analysis, and some density-based unsupervised methods have been introduced to identify outliers in low-density regions. However, these methods still suffer from inaccurate density estimation and limited capability in detecting diverse types of outliers. In this study, we propose a nonparametric outlier detection method with the fusion of density and distance (POD-FDD). The proposed method employs adaptive kernel density estimation based on natural neighborhoods, which reduces the sensitivity to parameters in density estimation. Moreover, the optimistic and pessimistic densities are introduced to enhance the reliability of density estimation in the local neighborhood. In addition, approximate reachability distance information is integrated to improve the capability of identifying cluster outliers. Ultimately, a robust parametric-free outlier detection method is developed to detect different types of outliers. Extensive comparative experiments and statistical significance analysis on synthetic and public datasets demonstrate its superior performance, achieving an average improvement of 1.97 % in the AUC metric.}, archive = {J_ASOC}, author = {Zhiyu Chen and Can Gao and Jie Zhou and Ying Yu}, doi = {10.1016/j.asoc.2025.113898}, journal = {Applied Soft Computing}, month = {12}, pages = {113898}, shortjournal = {Appl. Soft. Comput.}, title = {Fusing local density and approximate distance for nonparametric outlier detection}, volume = {185}, year = {2025}, }
• (2025). Prediction of the risk level of saltwater intrusion in the modaomen waterway of the pearl river estuary based on the deep temporal clustering model. ASOC, 185, 113897. (www)

  @article{ , abstract = {In recent years, extreme events caused by global climate change have intensified the phenomenon of saltwater intrusion (SWI) in estuaries. The nonlinear and non-stationary characteristics of estuarine SWI have led to an exponential decline in the timeliness of traditional regression prediction models, making it difficult to meet the operational needs of SWI forecasting. To address this, this study proposed a technical framework for SWI risk level forecasting based on temporal clustering, with its core innovation lying in algorithmic improvements for accurately characterizing complex disaster systems. The key challenges in forecasting SWI risk levels involved capturing the dynamic nonlinear relationships between multidimensional disaster factors (such as runoff, tide level, and wind) and SWI severity, as well as enhancing feature discriminability in label-limited scenarios. Accordingly, this study optimized algorithms through dual-path supervised and unsupervised learning: In the supervised learning framework, LightGBM, RF, XGBoost, and Extra trees were introduced as base learners into the Deep Forest (DF) model. The complementary feature-space partitioning of diverse learners was leveraged to improve the model’s ability to distinguish risk -level boundaries, achieving an average performance gain of 7.8 %. In the unsupervised learning framework, discriminative regularization was incorporated into the Extreme Learning Machine-Autoencoder (ELM-AE) model. By forcing features of samples from the same class to cluster toward the class center, the model’s feature separability for rare events (e.g., severe SWI) was enhanced, leading to an average performance improvement of 11 %. Finally, the optimal model was used to extract dynamic evolution patterns between multidimensional disaster factors and SWI risk levels, with interpretability analysis conducted for real-world forecasting. Notably, upstream flow sequences exhibited high distinguishability between no-SWI and severe-SWI, while mild and moderate SWI showed similar flow patterns, with tidal sequences being the primary differentiator. The algorithmic advancements not only enhanced the accuracy and efficiency of SWI forecasting but also provided a generalizable framework for risk classification in nonlinear hydrological systems.}, archive = {J_ASOC}, author = {Qingqing Tian and Hongyu Yang and Yu Tian and Peiyao Weng}, doi = {10.1016/j.asoc.2025.113897}, journal = {Applied Soft Computing}, month = {12}, pages = {113897}, shortjournal = {Appl. Soft. Comput.}, title = {Prediction of the risk level of saltwater intrusion in the modaomen waterway of the pearl river estuary based on the deep temporal clustering model}, volume = {185}, year = {2025}, }
• (2025). Contrastive learning with mirror-task for multimodal sentiment analysis. ASOC, 185, 113896. (www)

  @article{ , abstract = {Multi-Task Learning (MTL) in Multimodal Sentiment Analysis (MSA) involves implementing various parameter-sharing strategies among tasks. Currently, MSA primarily focuses on hard parameter-sharing mechanisms based on encoder sharing, while soft parameter-sharing is often neglected. To explore a reasonable combination of soft and hard mechanisms in MSA and optimize multimodal representations, along with multimodal contrastive learning, we propose D 3 MSA. It consists of D ouble network (primaryNet and MirrorNet), D ouble parameter-sharing strategies and D ouble contrastive learning modes for multimodal sentiment analysis. D 3 MSA utilizes hard-sharing to consolidate correlations between positive samples of intra-sample contrastive learning. In soft-sharing, we propose a pre-trained MirrorNet (MN) that generates negative samples by the learned inverse distributions. This optimizes the feature space of negative samples. MN interacts with the MSA task through soft-sharing during inter-sample contrastive learning. Experimental results demonstrate that our proposed method can achieve advanced performance on the CMU-MOSI and CMU-MOSEI datasets with lightweight training that requires only a small number of parameters.}, archive = {J_ASOC}, author = {Hang Shi and Lianmin Zhou and Yuanyuan Pu and Zhengpeng Zhao and Jinjing Gu and Dan Xu}, doi = {10.1016/j.asoc.2025.113896}, journal = {Applied Soft Computing}, month = {12}, pages = {113896}, shortjournal = {Appl. Soft. Comput.}, title = {Contrastive learning with mirror-task for multimodal sentiment analysis}, volume = {185}, year = {2025}, }
• (2025). A nearest-best neighbors optimization algorithm for constrained multimodal multiobjective problems. ASOC, 185, 113895. (www)

  @article{ , abstract = {Constrained multimodal multiobjective problems (CMMOPs) have multiple equivalent constrained Pareto optimal sets in the decision space corresponding to the identical constrained Pareto front in the objective space. The key to solving CMMOPs is how to balance feasibility, convergence, and diversity of solutions in both the decision and objective spaces. In view of this, this paper proposes a Nearest-Best neighbors optimization algorithm with constraint-based fitness (NBNOA) to solve CMMOPs. First, a constraint-based fitness assignment scheme is designed to assign specific fitness values to individuals in the population. Then, the Nearest-better-neighbor clustering method is adopted to identify the nearest-better neighbor and best neighbor of each individual according to the specific fitness values. On this basis, a Nearest-Best neighbors guided strategy is developed to guide the search direction of individuals, striking a better balance between exploration and exploitation capabilities. Moreover, a CDP-density elite selection mechanism is constructed to obtain feasible Pareto optimal solutions with higher precision and better diversity. Extensive experiments on two CMMOPs test suites demonstrated that the proposed NBNOA significantly outperforms nine state-of-the-art algorithms. Notably, NBNOA ranks first among all ten algorithms and achieves the best values for 23 out of 31 benchmark functions regarding the reciprocal of Pareto sets proximity and inverted generational distance. Furthermore, NBNOA is applied to a real-world CMMOP, verifying its effective practical application capability. Additionally, NBNOA is tested on two high-dimensional constrained multiobjective optimization problems test suites, further proving its competitive performance in solving complex problems.}, archive = {J_ASOC}, author = {Xuming Han and Ting Zhou and Limin Wang and Yali Chu}, doi = {10.1016/j.asoc.2025.113895}, journal = {Applied Soft Computing}, month = {12}, pages = {113895}, shortjournal = {Appl. Soft. Comput.}, title = {A nearest-best neighbors optimization algorithm for constrained multimodal multiobjective problems}, volume = {185}, year = {2025}, }
• (2025). Pathology characteristics-aware federated learning for weakly supervised nuclei segmentation. ASOC, 185, 113894. (www)

  @article{ , abstract = {In federated learning, a key challenge in nuclei segmentation lies in data heterogeneity, primarily resulting from the diverse sources of pathology images. Nuclei in pathological images are typically small and densely distributed, making accurate annotation highly labor-intensive and reliant on specialized expertise. Weakly supervised learning is widely adopted for this task, as it only requires point annotations at the centers of nuclei. However, point annotations lack precise boundary information, thereby exacerbating the difficulties introduced by data heterogeneity. To address this issue, we propose a preprocessing strategy that leverages the unique optical properties of H&E stained images to generate Contrast-Difference Enhanced Images (CDEI). These CDEI highlight nucleus boundaries to varying extents based on the tonal characteristics of H&E stained images. Building on this strategy, we propose a Multi-source Hierarchical Enhancement Network (MHEN) as the client-side architecture. MHEN takes both the H&E stained images and the corresponding CDEI as input, effectively mitigating the limitations of weak labels by incorporating enhanced boundary cues. Furthermore, considering the characteristics of nuclei segmentation, we design a Federated Nuclei-Weighted Aggregation strategy on the server side. This strategy estimates each client’s contribution to the global model by quantifying the number of nuclei present in its local pathology images. To thoroughly assess the effectiveness of our approach, we compare it with both conventional weakly supervised methods and existing federated weak supervision frameworks. The experimental results demonstrate that our proposed federated learning framework for weakly supervised nuclei segmentation significantly outperforms existing methods. Our source code is available on GitHub. 1}, archive = {J_ASOC}, author = {Yi Qian and Xipeng Pan and Yimin Wen and Xinjun Bian and Shilong Song}, doi = {10.1016/j.asoc.2025.113894}, journal = {Applied Soft Computing}, month = {12}, pages = {113894}, shortjournal = {Appl. Soft. Comput.}, title = {Pathology characteristics-aware federated learning for weakly supervised nuclei segmentation}, volume = {185}, year = {2025}, }
• (2025). Frequency decomposition and patch modeling framework for time-series forecasting. ASOC, 185, 113890. (www)

  @article{ , abstract = {Time series forecasting is widely applied across diverse fields, including finance, transportation, and energy, and has made significant contributions in these areas. However, in real-world applications, time series data can be complex and dynamic. Current methodologies still encounter several challenges in managing high-dimensional data, extracting intricate features, and making long-term forecasts. In this study, we propose a Frequency Decomposition and Patch Modeling Framework (FPF). Our FPF consists of the Frequency Domain Decomposition Block (FDB) and the Dual Patch Modeling Block (DPMB). DPMB consists of Patch Enhancement Block and Patch Mixing Block. First, FDB transforms the input sequence to the frequency domain through the Fast Fourier Transform and designs frequency masks to decompose the data into high-frequency and low-frequency components, to extract fast-changing patterns and trend information respectively. Subsequently, DPMB divides the components into patches, where the high-frequency components are modeled by MLP-based Patch Enhancement Block to capture local features, and the low-frequency components are modeled by Transformer-based Patch Mixing Block to capture global dependencies and cross-patch correlations. We conducted comprehensive experiments using seven real-world time series forecasting datasets, including ETT, Traffic, Electricity, and Weather. The findings indicate that this method demonstrates superior performance in the field of time series forecasting.}, archive = {J_ASOC}, author = {Denghui Xu and Hua Wang and Fan Zhang}, doi = {10.1016/j.asoc.2025.113890}, journal = {Applied Soft Computing}, month = {12}, pages = {113890}, shortjournal = {Appl. Soft. Comput.}, title = {Frequency decomposition and patch modeling framework for time-series forecasting}, volume = {185}, year = {2025}, }
• (2025). OdML: An optimization-driven meta-learning framework for multi-task approximate model predictive control. ASOC, 185, 113882. (www)

  @article{ , abstract = {Recent advances in approximate model predictive control (MPC) have leveraged deep neural networks (DNNs) to imitate optimal control laws, significantly reducing the computational cost of real-time optimization. However, these methods often struggle to adapt efficiently to previously unseen system dynamics and suffer from poor data efficiency in real time. In this paper, we address these challenges by treating control problems with different system parameters as separate tasks within a meta-learning scope. Under this scope, we propose OdML, an Optimization-driven Meta-Learning approach for approximate MPC that enables rapid online fine-tuning of DNN-based controllers. We introduce an optimization-driven fine-tuning mechanism that allows for fast adaptation to new tasks using limited online data, without requiring full knowledge of the system parameters. Furthermore, to ensure safety during adaptation, we incorporate control barrier functions into the optimization process, allowing the controller to satisfy safety constraints even under previously unseen conditions. We demonstrate the effectiveness of OdML through various simulation scenarios, highlighting its ability to achieve safe and efficient control.}, archive = {J_ASOC}, author = {Junbo Tong and Shuhan Du and Daming Shi and Wenhui Fan}, doi = {10.1016/j.asoc.2025.113882}, journal = {Applied Soft Computing}, month = {12}, pages = {113882}, shortjournal = {Appl. Soft. Comput.}, title = {OdML: An optimization-driven meta-learning framework for multi-task approximate model predictive control}, volume = {185}, year = {2025}, }
• (2025). A deep reinforcement learning trader without offline training. ASOC, 185, 113881. (www)

  @article{ , abstract = {In this paper we pursue the question of a fully online trading algorithm (i.e. one that does not need offline training on previously gathered data). For this task we consider Double Deep Q -learning in the episodic setting with Fast Learning Networks approximating the expected reward Q . Additionally, we define the possible terminal states of an episode in such a way as to introduce a mechanism to conserve some of the money in the trading pool when market conditions are seen as unfavourable. Some of these money are taken as profit and some are reused at a later time according to certain criteria. After describing the algorithm, we test it using 1-minute-tick price data for 4 major cryptocurrencies from Binance. We see that the agent performs better than trading with randomly chosen actions on each timestep. And it does so when tested on the whole dataset for a given market as well as on different subsets, representing different market trends.}, archive = {J_ASOC}, author = {Boian Lazov}, doi = {10.1016/j.asoc.2025.113881}, journal = {Applied Soft Computing}, month = {12}, pages = {113881}, shortjournal = {Appl. Soft. Comput.}, title = {A deep reinforcement learning trader without offline training}, volume = {185}, year = {2025}, }
• (2025). Density peak clustering algorithm based on boundary elimination and backbone construction. ASOC, 185, 113880. (www)

  @article{ , abstract = {Density peak clustering (DPC) is an effective clustering algorithm, but it still has some problems and faces some challenges. For instance, it cannot identify the variable density datasets, the assignment strategy is easy to produce domino phenomenon, and the clustering results of DPC and its improved algorithms are easily affected by the intersection points between clusters. To solve these problems, in this paper, we propose a novel density peak clustering algorithm based on boundary elimination and backbone construction, called BEBC-DPC. A new local density is defined based on the natural neighbor search, and the boundary degree is defined by the position relationship between each point and its neighbors, which accurately describes the local distribution information of the point. The boundary points of clusters are eliminated by fusing the density and the boundary degree, which reduces the influence of the intersection points on the cluster division. In addition, the cluster backbone construction method based on representative points and representative sets is proposed. The density relationship among non-boundary points is used to form representative sets, and the similarity between representative sets is used to construct the cluster backbones, which can effectively describe the overall distribution structure characteristics of the clusters. Moreover, the adjacency degree of each boundary point is defined by using the neighbor information and distance information, and the boundary points are gradually assigned to the most appropriate cluster backbone based on it to complete the clustering. Finally, sufficient experiments are performed on synthetic, UCI and image datasets, and the proposed BEBC-DPC is compared with DPC and its improved algorithms. Experimental results show the effectiveness of the proposed BEBC-DPC on various types of datasets.}, archive = {J_ASOC}, author = {Zhizhong Zhao and Sugen Chen and Cong Hu}, doi = {10.1016/j.asoc.2025.113880}, journal = {Applied Soft Computing}, month = {12}, pages = {113880}, shortjournal = {Appl. Soft. Comput.}, title = {Density peak clustering algorithm based on boundary elimination and backbone construction}, volume = {185}, year = {2025}, }
• (2025). Multi-criteria decision analysis-based framework for supply chain management evaluation with multi-dimensional sensitivity analysis: A green logistics perspective. ASOC, 185, 113879. (www)

  @article{ , abstract = {The transition toward sustainable Supply Chain Management (SCM) presents complex decision-making challenges that require robust and transparent evaluation methods. Effective decision-making in this context requires robust evaluation methods capable of handling complex, multi-dimensional aspects of the decision process. This study proposes a comprehensive decision-support framework that integrates Multi-Criteria Decision Analysis (MCDA) with advanced sensitivity analysis techniques to assess the stability and reliability of alternative rankings under uncertainty. The framework combines established MCDA methods with submodel exclusion analysis and the Comprehensive Sensitivity Analysis Method (COMSAM), which allows for systematic perturbation of the decision matrix and evaluation of robustness across multiple dimensions. Applied to a green logistics case study, the framework demonstrates its capacity to support sustainable decision-making by accounting for varying stakeholder priorities and data uncertainty. The approach offers methodological advances in robustness assessment, practical relevance for sustainability-focused SCM, and open-source implementation to ensure reproducibility. The proposed framework is adaptable to a wide range of decision contexts, contributing to the development of more reliable and explainable MCDA-based evaluations.}, archive = {J_ASOC}, author = {Jakub Więckowski and Wojciech Sałabun}, doi = {10.1016/j.asoc.2025.113879}, journal = {Applied Soft Computing}, month = {12}, pages = {113879}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-criteria decision analysis-based framework for supply chain management evaluation with multi-dimensional sensitivity analysis: A green logistics perspective}, volume = {185}, year = {2025}, }
• (2025). Hierarchical layered multigraph network with scale importance estimation for image classification. ASOC, 185, 113877. (www)

  @article{ , abstract = {This work introduces a novel image representation and processing approach using Graph Neural Networks (GNNs). We propose a multigraph representation named H i E rarchical L ayered M ultigraph (HELM), which explicitly encodes spatial and hierarchical relationships as distinct edge types, overcoming the limitations of existing methods that fail to fully exploit relational information in images. A multi-scale representation is generated through hierarchical segmentation of a superpixel base graph, enabling the computation of spatial and hierarchical relationships within and across scales. To effectively process this multi-relational information, we introduce the H i E rarchical L ayered M ultigraph Net work (HELMNet), a novel GNN architecture incorporating specialized mechanisms for selectively aggregating and fusing information from each distinct edge type. Additionally, it includes a Region Graph Readout (RGR) module that employs an attention mechanism to dynamically weight the contribution of each representation scale during the aggregation process for classification. Experimental results demonstrate the greater efficacy of HELMNet for image classification. Compared to hierarchical models that do not distinguish between edge types, HELMNet obtains substantial average accuracy gains of 2.1% (significant at 3% level) and 10% (significant at 0.1% level) on the CIFAR-10 and STL-10 datasets, respectively. On the EUROSAT dataset, HELMNet achieves over 95% accuracy, requiring only 0.73% of the best-performing state-of-the-art model size (in number of parameters). For the more demanding and high-resolution RESISC45 dataset, the proposed model still delivers impressive results, achieving an accuracy of over 85%.}, archive = {J_ASOC}, author = {João Pedro Oliveira Batisteli and Nicolas Passat and Silvio Jamil Ferzoli Guimarães and Zenilton Kleber Gonçalves do Patrocínio Júnior}, doi = {10.1016/j.asoc.2025.113877}, journal = {Applied Soft Computing}, month = {12}, pages = {113877}, shortjournal = {Appl. Soft. Comput.}, title = {Hierarchical layered multigraph network with scale importance estimation for image classification}, volume = {185}, year = {2025}, }
• (2025). Boosting fuzzy classification rules under footprint of uncertainty of type-2 fuzzy sets. ASOC, 185, 113876. (www)

  @article{ , abstract = {This paper presents a novel approach to improving classification accuracy in fuzzy rule-based systems by selecting Embedded Type-1 Fuzzy Sets (ET1 FSs) from Footprint of Uncertainty (FOU) areas. The method consists of three stages: (1) learning Type-1 fuzzy rules from predefined linguistic variables, (2) generating FOU areas with Interval Type-2 Fuzzy Sets (IT2 FSs), and (3) using the adaboost ensemble method, where multi-class problems are decomposed into binary classification tasks, and ET1 FSs are iteratively selected via genetic algorithms. By relying on IT2 FSs for the flexible partitioning of classification boundaries, the method enhances accuracy while addressing challenges in high-dimensional and multi-class problems. Experiments were performed on 12 UCI datasets and three image classification tasks using features from pre-trained convolutional neural networks. These datasets were selected to ensure diversity in dimensionality and class distribution. Comparative analyses with several state-of-the-art classification methods demonstrate that IT2 FSs can be effectively used to develop accurate classification systems. Additionally, this work analyzes trade-offs between complexity and interpretability by tuning FOU size (distinguishability) and adjusting the number of rules. The results show that a balanced FOU size and rule count yield better accuracy gains than either alone. Furthermore, several suitable trade-off regions with their parameters are presented.}, archive = {J_ASOC}, author = {Jidong Li and Jian Cui and Qian Su}, doi = {10.1016/j.asoc.2025.113876}, journal = {Applied Soft Computing}, month = {12}, pages = {113876}, shortjournal = {Appl. Soft. Comput.}, title = {Boosting fuzzy classification rules under footprint of uncertainty of type-2 fuzzy sets}, volume = {185}, year = {2025}, }
• (2025). Reliable type 2 fuzzy Min–Max neural networks for pattern classification. ASOC, 185, 113875. (www)

  @article{ , abstract = {The Fuzzy Min–Max (FMM) algorithm is a powerful classification method capable of handling non-linear class boundaries and making both hard and soft decisions while learning from online data. However, it faces significant challenges, including sensitivity to the expansion coefficient, information loss during the contraction stage, and the overlap problem. To address these limitations, we propose a Reliable Type-2 Fuzzy Min–Max (RT2FMM) algorithm, which incorporates type-2 fuzzy logic to consider hyperbox uncertainty and effectively resolve the overlap problem. By assigning distinct certainties to overlapping regions, RT2FMM eliminates the need for the contraction stage and the overlap test. Additionally, we introduce weighted factors for hyperboxes, which enhances the reliability of membership values and models their mutual effects. Our comprehensive experimental evaluation across twenty datasets demonstrates that RT2FMM significantly outperforms existing FMM-based models in terms of robustness and accuracy. The Friedman test further confirms the superior performance of RT2FMM compared to commonly used classifiers, highlighting its potential as a robust solution for complex classification tasks.}, archive = {J_ASOC}, author = {Ali Nik-Khorasani and Mohammad-R. Akbarzadeh-T.}, doi = {10.1016/j.asoc.2025.113875}, journal = {Applied Soft Computing}, month = {12}, pages = {113875}, shortjournal = {Appl. Soft. Comput.}, title = {Reliable type 2 fuzzy Min–Max neural networks for pattern classification}, volume = {185}, year = {2025}, }
• (2025). Fortifying vision models: A comprehensive survey of defences against adversarial examples. ASOC, 185, 113874. (www)

  @article{ , abstract = {Artificial Intelligence and Machine learning (ML) have seen many advancements in the past two decades. It has led to the creation of several techniques, including Deep Neural Networks (DNN), Convolution Neural Networks (CNN), Autoencoders, Generative Adversarial Networks (GAN) and Diffusion models. These techniques have been applied to various real-world applications, such as self-driving cars, medical diagnosis and voice assistants. Despite these advancements, a carefully crafted input can fool the ML model. Such attacks are known as adversarial examples. It is a serious threat to safety critical systems. This survey provides a comprehensive review of defences against adversarial examples by tracing their evolution from early empirical methods to more principled, theoretically grounded approaches. We systematically categorise defences based on their underlying mechanisms. In addition to surveying state-of-the-art techniques, we spotlight emerging trends such as generative defences and diffusion-based purification. Finally, we identify persistent vulnerabilities and outline promising directions for future research towards building truly resilient vision models. This work aims to equip researchers and practitioners with a deep understanding of current defences and inspire innovation in adversarial robustness for the next generation of vision applications.}, archive = {J_ASOC}, author = {Siddheshwar Kumar and Shashank Srivastava and Shashwati Banerjea}, doi = {10.1016/j.asoc.2025.113874}, journal = {Applied Soft Computing}, month = {12}, pages = {113874}, shortjournal = {Appl. Soft. Comput.}, title = {Fortifying vision models: A comprehensive survey of defences against adversarial examples}, volume = {185}, year = {2025}, }
• (2025). AOT-PixelNet: Lightweight and interpretable detection of forged images via adaptive orthogonal transform. ASOC, 185, 113873. (www)

  @article{ , abstract = {Generative image detection faces persistent challenges in terms of generalization and interpretability, limiting its reliability in complex scenarios. To address these issues, we propose AOT-PixelNet, a lightweight and interpretable detection framework that integrates an Adaptive Orthogonal Transform (AOT) module with a streamlined 1 × 1 convolution-based PixelNet architecture. The AOT module leverages diverse orthogonal transforms, such as FFT and DCT, to extract informative frequency-domain features, thereby enhancing sensitivity to medium- and high-frequency artifacts. Meanwhile, PixelNet minimizes parameter count (only 0.98 million) while effectively capturing cross-channel inconsistencies and mitigating overfitting. Experimental evaluations on multiple unseen GAN and diffusion-based datasets demonstrate that AOT-PixelNet achieves superior performance with minimal computational cost. Specifically, it outperforms the NPR method by 0.6% and 11.76% on the ForenSynths and GenImage datasets, respectively, validating the framework’s robustness, effectiveness, and interpretability.}, archive = {J_ASOC}, author = {Dengtai Tan and Deyi Yang and Boao Tan and Chengyu Niu and Yang Yang and Shichao Li}, doi = {10.1016/j.asoc.2025.113873}, journal = {Applied Soft Computing}, month = {12}, pages = {113873}, shortjournal = {Appl. Soft. Comput.}, title = {AOT-PixelNet: Lightweight and interpretable detection of forged images via adaptive orthogonal transform}, volume = {185}, year = {2025}, }
• (2025). Status prediction using attentive and cascaded deep network by fault tolerant and priority-based scheduling for load balancing in cloud. ASOC, 185, 113872. (www)

  @article{ , abstract = {The cloud facilitates the user to complete their work utilizing the cost strategy of pay-as-you-go, which is based on the consumed Virtual Machine (VM) hours. Thus, the scheduler must offer the highest throughput to attain efficient allocation of resources in the cloud paradigm. Cloud services are dependent on characteristics such as fault tolerance, security, scalability, and availability. Hence, an effective scheduler is necessary to arrange the scheduling tasks and adjust the server loads. Typically, a load-balancing task focuses on detecting the overloaded and under-loaded nodes and adjusting the load between them. When considering the significant role of fault-tolerance in load-balancing algorithms, it seems to suffer from poor organization and a lack of in-depth experiments in this sector. This paper proposes a new task for the load-balancing operation. Initially, task scheduling is performed where the fault tolerance and the priority-aided scheduling approach are adopted. Furthermore, resource optimization is carried out in the scheduling task using Randomly Improved Electric Fish Optimization (RIEFO). To validate the load balancing operation, several multi-objective functions such as resource utilization, delay, time, makespan, active servers, throughput, success rate, fault tolerance rate, and energy consumption are derived. Moreover, because of the system’s dynamic environment, the status of the server varies simultaneously. The server status prediction is significant in allocating the tasks to the server or the VM resources. Thus, the Attention-based Cascaded Residual Bidirectional Long Short-Term Memory (ACRes-BiLSTM) is employed to predict the server status before performing the resource allocation. Finally, the tasks are scheduled effectively using the predicted server status. The performance is estimated using numerous performance metrics.}, archive = {J_ASOC}, author = {Gudivada Lokesh and Kasarapu Ramani and K.K. Baseer}, doi = {10.1016/j.asoc.2025.113872}, journal = {Applied Soft Computing}, month = {12}, pages = {113872}, shortjournal = {Appl. Soft. Comput.}, title = {Status prediction using attentive and cascaded deep network by fault tolerant and priority-based scheduling for load balancing in cloud}, volume = {185}, year = {2025}, }
• (2025). A hybrid deep learning framework for detecting grape leaf diseases by integrating ResNeSt-50 and CBAM attention into DeepLabv3 +. ASOC, 185, 113871. (www)

  @article{ , abstract = {Grape leaf diseases, such as black rot, powdery mildew, and downy mildew, pose a significant threat to global viticulture, leading to substantial yield losses and reduced fruit quality. Early and accurate identification of these diseases is essential for precision agriculture and sustainable crop management. This study presents a comprehensive comparison of traditional and deep learning-based image segmentation methods for detecting grape leaf lesions. A series of classical segmentation techniques, including Mean Shift, Fuzzy C-Means (FCM), Normalized Cut, K-Means, and Fuzzy K-Means (FKM), were evaluated alongside an advanced DeepLabv3 + model. The baseline DeepLabv3 + architecture was further enhanced by integrating a ResNeSt-50 backbone with various attention mechanisms, including Squeeze-and-Excitation (SE) Block, Convolutional Block Attention Module (CBAM), Bottleneck Attention Module (BAM), Self-Attention, and Dual Attention Network (DANet). Among all models, DeepLabv3 + with ResNeSt-50 and CBAM achieved the highest performance, attaining 98.2 % accuracy, 97.1 % precision, 96.7 % recall, 96.6 % mean Intersection over Union (mIoU), and a 96.8 % Dice Score. The results demonstrate that attention-augmented deep networks significantly outperform classical methods, especially in handling complex lesion structures under diverse environmental conditions. While traditional algorithms remain useful in resource-constrained scenarios, deep learning models, particularly those enhanced with spatial and channel-wise attention, offer greater accuracy and robustness, making them ideal for integration into intelligent agricultural platforms such as drones, mobile scanners, and automated disease monitoring systems. Future work will focus on incorporating temporal and multimodal data, expanding dataset diversity, and optimizing lightweight models for real-time deployment on edge devices.}, archive = {J_ASOC}, author = {Kittipol Wisaeng}, doi = {10.1016/j.asoc.2025.113871}, journal = {Applied Soft Computing}, month = {12}, pages = {113871}, shortjournal = {Appl. Soft. Comput.}, title = {A hybrid deep learning framework for detecting grape leaf diseases by integrating ResNeSt-50 and CBAM attention into DeepLabv3 +}, volume = {185}, year = {2025}, }
• (2025). A stochastic proportional–differential TLBO algorithm and its applications for wafer transfer finger. ASOC, 185, 113870. (www)

  @article{ , abstract = {The Teaching-Learning-Based optimization (TLBO) algorithm, which includes teacher phase and learner phase, is a widely used method for global optimization. However, TLBO will experience premature convergence and get stuck in local optimum when faced with complex optimization challenges. Especially when tackling complex problems in practical engineering applications, which involves multiple variables and numerous constraints. To address this issue, a new variant termed Stochastic Proportional–Differential TLBO (SPD-TLBO) has been developed. The SPD phase allows students to learn not only from the current population but also from previous stochastic errors and their generation differences using adaptive random operators. By incorporating an SPD operator into the original TLBO framework, the algorithm’s search diversity is enhanced, reducing the likelihood of premature convergence to local optimum. The experimental results conducted at the IEEE Conference on Evolutionary Computation 2014 (CEC 2014) indicated that the proposed SPD-TLBO algorithm achieved an effective balance between exploration and exploitation capabilities. Specifically, the SPD-TLBO algorithm achieves the highest ranking in 21 out of 30 cases (70%) for 30-dimensional problems and 18 out of 30 cases (60%) for 50-dimensional problems. Statistical tests and convergence analyses show that the SPD-TLBO algorithm outperforms other algorithms in solving global optimization problems. Additionally, when applied to engineering optimization problems, the SPD-TLBO algorithm shows significant advantages over other algorithms. Therefore, the SPD-TLBO algorithm is further applied to optimize the structure of a wafer transfer finger in semiconductor manufacturing.}, archive = {J_ASOC}, author = {Jinfeng Sun and Yunlang Xu and Haibo Zhou}, doi = {10.1016/j.asoc.2025.113870}, journal = {Applied Soft Computing}, month = {12}, pages = {113870}, shortjournal = {Appl. Soft. Comput.}, title = {A stochastic proportional–differential TLBO algorithm and its applications for wafer transfer finger}, volume = {185}, year = {2025}, }
• (2025). Prediction of the health effects of COVID using hybrid classifier with attention-based LSTM-deep temporal convolution network. ASOC, 185, 113867. (www)

  @article{ , abstract = {Coronavirus Disease 2019 (COVID-19) is an infectious illness that affects both humans and animals. Individuals infected with COVID-19 are prone to lung complications during the recovery phase . Radiography and Computed Tomography (CT) are the most commonly used methods for diagnosing lung-related diseases. The primary aim of this paper is to assess the impact of COVID-19 on patients’ lungs, heart, and blood sugar levels using a deep learning-based approach. Initially, data related to the heart, blood sugar levels, and lungs of COVID-19-infected individuals are collected. From this dataset, three types of features are extracted. Deep features are obtained using Iterated Dilated Convolutional Neural Networks (IDCNN). From these deep features, which are obtained from the IDCNN, the optimal weighted features are derived by implementing the Hybrid Dolphin Pod Cuttlefish Optimization (HDPCO) algorithm. Subsequently, the HDPCO algorithm is also employed for optimal feature extraction. In addition, dimensionality reduction is performed using Principal Component Analysis (PCA). These three sets of features from the IDCNN, HDPCO, and PCA, are then fused into a single feature set . This fused feature set is fed into a hybrid classifier composed of a Deep Temporal Convolutional Network (DTCN) and an Attention-based Long Short-Term Memory (ALSTM) network . The classifier parameters are optimized using the HDPCO algorithm. The output from the hybrid classifier provides the final prediction result. Experimental results demonstrate that the proposed COVID-19 impact prediction model significantly outperforms existing models in terms of prediction accuracy and efficiency.}, archive = {J_ASOC}, author = {Sadanandam Kalvala and B. Baranidharan}, doi = {10.1016/j.asoc.2025.113867}, journal = {Applied Soft Computing}, month = {12}, pages = {113867}, shortjournal = {Appl. Soft. Comput.}, title = {Prediction of the health effects of COVID using hybrid classifier with attention-based LSTM-deep temporal convolution network}, volume = {185}, year = {2025}, }
• (2025). Quantum fuzzy logic for edge detection: A demonstration on NISQ hardware. ASOC, 185, 113866. (www)

  @article{ , abstract = {Quantum computing offers the potential to enhance computational efficiency beyond classical methods, but practical implementation remains challenging due to the limitations of Noisy Intermediate-Scale Quantum (NISQ) hardware, namely, restricted qubit counts, limited connectivity, and the presence of noise and decoherence. This study presents a novel approach to edge detection by leveraging a recently developed Quantum Fuzzy Inference Engine, implemented on a NISQ device. We introduce an optimized quantum circuit for its implementation, reducing qubit requirements and gate depth to improve execution on NISQ hardware. To overcome constraints related to large-scale image processing, a hybrid quantum–classical lookup table approach is employed. Edge detection performance is evaluated on the Berkeley Segmentation Data Set and Benchmarks 500 dataset under different conditions, including classical execution, ideal quantum simulation, noisy quantum simulation, and NISQ hardware calculation. Results demonstrate that the quantum fuzzy logic-based edge detection achieves outcomes comparable to classical methods by using fewer operations, marking a step toward practical quantum-enhanced image processing.}, archive = {J_ASOC}, author = {G. Nunziata and S. Crisci and G. De Gregorio and R. Schiattarella and G. Acampora and L. Coraggio and N. Itaco}, doi = {10.1016/j.asoc.2025.113866}, journal = {Applied Soft Computing}, month = {12}, pages = {113866}, shortjournal = {Appl. Soft. Comput.}, title = {Quantum fuzzy logic for edge detection: A demonstration on NISQ hardware}, volume = {185}, year = {2025}, }
• (2025). Multi-agent cooperation-based bi-criteria evolutionary many-objective optimization. ASOC, 185, 113865. (www)

  @article{ , abstract = {Many-objective evolutionary algorithms (MaOEAs) excel in solving many-objective optimization problems (MaOPs), which are mainly classified into two frameworks: the Pareto domination and the non-Pareto domination. The Pareto criterion (PC) obtains a well-converged solution set in multi-objective spaces through the Pareto dominance relationship between solutions. However, insufficient environmental selection pressure in many-objective spaces leads to slow convergence. The non-Pareto criterion (NPC) enhances the selection pressure by evaluating the solution set with a set of sortable scalar values. However, it is difficult to ensure the Pareto-optimal consistency of convergence and distribution when facing highly irregular Pareto fronts (PFs). Therefore, combining the two sets of criteria can satisfy the demand for uniform distribution while bringing significant selection pressure. A multi-agent cooperative strategy is proposed in this study to realize the combination of the two criteria. This strategy controls the evolutionary direction of two populations separately by deploying two agents, and promotes cooperative evolution between these populations through the exchange and flow of large amounts of information. In order to better realize the cooperative effect, we adopt the multi-agent reinforcement learning (MARL) strategy to accurately regulate the variation operator and parameter configurations of the bi-population. In addition, the effectiveness of the proposed method is validated on 74 test problems (DTLZ, WFG, and UF) and 3 real-world problems. The results show that the proposed algorithm is more competitive than 6 state-of-the-art algorithms.}, archive = {J_ASOC}, author = {Jiazheng Li and Yuan Liu and Juan Zou and Shuyi Liu and Shengxiang Yang and Jinhua Zheng}, doi = {10.1016/j.asoc.2025.113865}, journal = {Applied Soft Computing}, month = {12}, pages = {113865}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-agent cooperation-based bi-criteria evolutionary many-objective optimization}, volume = {185}, year = {2025}, }
• (2025). ORESTE methodology within a circular intuitionistic fuzzy framework for preferential outranking in hybrid cloud service selection. ASOC, 185, 113864. (www)

  @article{ , abstract = {This paper advances the ORESTE (Organísation, Rangement Et Synthèse de Données Relarionnelles) methodology within the Circular Intuitionistic Fuzzy (CIF) framework, highlighting its potential in practical decision analytics. The study first enhances CIF aggregation by employing the generalized mean technique, offering a flexible way to combine evaluative ratings and significance weights. Through modulation of the averaging parameter, decision-makers are able to accentuate either lower or higher values, thereby overcoming the constraints associated with conventional arithmetic means. The framework further improves decision precision through CIF similarity-driven appraisal indices, which utilize refined similarity metrics grounded in axiomatic properties such as symmetry, boundedness, identity, and monotonicity. These indices quantify the similarity between evaluative ratings and anchor references, while also revealing indifference and incomparability—thus equipping decision-makers with a comprehensive toolset for handling uncertainty. The CIF ORESTE framework comprises two methodologies. CIF ORESTE I delivers a global weak ranking using similarity-driven indices and generalized projection-related distances. CIF ORESTE II addresses the limitations of weak rankings by incorporating an Indifference-Preference-Incomparability (I-P-R) structure, which uses mean and net preference intensities to establish thresholds and clarify outranking relations. Applied to the evaluation of hybrid cloud services for a technology corporation, the CIF ORESTE framework demonstrates its effectiveness in resolving group decisions, managing uncertainty, and structuring preferences. Comparative analyses further underscore its robustness in handling CIF-based data and delivering reliable results.}, archive = {J_ASOC}, author = {Ting-Yu Chen}, doi = {10.1016/j.asoc.2025.113864}, journal = {Applied Soft Computing}, month = {12}, pages = {113864}, shortjournal = {Appl. Soft. Comput.}, title = {ORESTE methodology within a circular intuitionistic fuzzy framework for preferential outranking in hybrid cloud service selection}, volume = {185}, year = {2025}, }
• (2025). Prompted complex context generation guided fine-grained ship recognition. ASOC, 185, 113856. (www)

  @article{ , abstract = {Fine-grained ship recognition in complex marine environments is challenged by background interference, high inter-class similarity, and limited labeled data. Existing methods often rely on inefficient cascades or holistic feature extraction, which limits both accuracy and efficiency. To address these issues, we propose a Prompted Complex Context Generation Guided Fine-Grained Ship Recognition framework, consisting of two core modules. The Cross-Attention Context Generation Module utilizes a diffusion model to generate diverse background images from prompts, maintaining target consistency and enriching the training data to mitigate data scarcity. It also employs a cross-attention map to highlight target-relevant regions, guiding the Attention Map Guided Fusion Module. The Attention Map Guided Fusion Module adopts a dual-branch transformer architecture: one branch extracts global features from background-enhanced images, and the other captures local features through attention-guided cropping of target-specific regions. By integrating both global and local features, our method effectively identifies key target characteristics. Experimental results demonstrate that our approach achieves 97.04% accuracy on the publicly available MAR-ships dataset and 84.57% accuracy on the challenging GCS dataset, outperforming state-of-the-art methods.}, archive = {J_ASOC}, author = {Runtian Wang and Kejun Wu and Renjie Qiao and Chunsheng Yang and Chengtao Cai}, doi = {10.1016/j.asoc.2025.113856}, journal = {Applied Soft Computing}, month = {12}, pages = {113856}, shortjournal = {Appl. Soft. Comput.}, title = {Prompted complex context generation guided fine-grained ship recognition}, volume = {185}, year = {2025}, }
• (2025). Genetic algorithm for path-based testing of component outage situations in IoT system processes. ASOC, 185, 113854. (www)

  @article{ , abstract = {Component outages often affect IoT system operations and processes. These components can be physical devices, infrastructure parts, or system modules. Among other possible causes, outages are often due to limited or intermittent network connectivity. To ensure reliable operations, connection outage scenarios must be reviewed systematically, which is especially important for critical systems. Path-based testing techniques are preferable for this task, as they sequence events in the system and, therefore, allow to verify the effects of the limited network connectivity on the system processes. Because the available path-based testing techniques provide only a limited ability to solve this problem effectively, in this study, we propose an adaptation of a genetic algorithm to generate specialized test paths from a model that captures the system under test processes. Compared with the four path-based testing alternatives for solving the testing problem, the proposed algorithm yielded the best results in all four defined test set metrics for the two defined test coverage criteria. Regarding the average total length of the test paths, which served as a proxy for testing costs, those produced by the proposed adapted genetic algorithm outperformed the best of the proposed baselines by 23.5% and 29% for individual test coverage criteria.}, archive = {J_ASOC}, author = {Matej Klima and Miroslav Bures and Bestoun S. Ahmed and Hannan Hindy and Xavier Bellekens and Angelo Gargantini}, doi = {10.1016/j.asoc.2025.113854}, journal = {Applied Soft Computing}, month = {12}, pages = {113854}, shortjournal = {Appl. Soft. Comput.}, title = {Genetic algorithm for path-based testing of component outage situations in IoT system processes}, volume = {185}, year = {2025}, }
• (2025). Multi-label feature selection via asymmetric redundancy and variable precision dependency. ASOC, 185, 113852. (www)

  @article{ , abstract = {Multi-label feature selection is an effective data preprocessing technique that can significantly mitigate the challenges posed by high-dimensional features in multi-label learning. However, the exploration of feature-label correlations has often been strictly limited to inclusion relationships, while ignoring the fusion of local and global label information. Moreover, most previous work has typically assumed that redundancy between features is fully symmetric, overlooking the valuable insights that asymmetric redundancy provides for designing feature selection. To address these issues, this paper proposes a novel multi-label feature selection via asymmetric redundancy and variable precision dependency. Specifically, it constructs a conditional probability model to reflect the local label semantics, incorporating this into the construction of the variable precision dependency through a fusion indicator. Subsequently, the optimistic and pessimistic information overlap between features is discussed, allowing variable precision granularity to capture asymmetric redundancy between features. Building upon this, an information fusion method is proposed to quantify the pessimistic asymmetric redundancy between features by inducing knowledge granularity in the feature space. Finally, a comprehensive evaluation metric, Maximum Correlation-maximum Discrimination-minimum Redundancy (MCDR), is proposed to evaluate the significance of features. The experimental results on fifteen multi-label benchmark datasets indicate that the proposed method outperforms the other seven state-of-the-art methods.}, archive = {J_ASOC}, author = {Wenbin Qian and Xiwen Lu and Shiming Dai and Jintao Huang}, doi = {10.1016/j.asoc.2025.113852}, journal = {Applied Soft Computing}, month = {12}, pages = {113852}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-label feature selection via asymmetric redundancy and variable precision dependency}, volume = {185}, year = {2025}, }
• (2025). A multi-granularity decision tree algorithm based on variable precision rough sets and zentropy. ASOC, 185, 113851. (www)

  @article{ , abstract = {The existing decision tree algorithms often use a single-layer measure to process data, which cannot fully consider the complex interactions and dependencies between different granularity levels. In addition, decision tree algorithms inevitably face the issue of multi-value preference, which may lead to the selection of unreasonable attributes in the process of partition, thereby affecting the performance of the algorithms. Therefore, this paper proposes an improved decision tree algorithm, called Ze-VNDT, which combines variable precision rough sets with Zentropy. First, to avoid the information loss caused by data discretization, this paper introduces variable precision neighborhood rough sets for data processing. Second, by analyzing the granularity level structure within the variable precision neighborhood rough set model, knowledge uncertainty is analyzed from three granularity levels: decision classes, approximate relations, and similarity classes. We describe the uncertain knowledge from the overall to the internal using the idea of going from coarse to fine, and design a Zentropy to measure uncertainty. To address the issue of multi-value preference, an adaptive weighted Zentropy uncertainty measure is designed based on the definition of uncertainty measure based on Zentropy. Third, when constructing the improved decision tree algorithm, the optimal attributes are selected based on the designed uncertainty measure. Finally, numerical experiments on 18 UCI datasets validated the effectiveness and rationality of the proposed algorithm. The experimental results showed that, compared to traditional algorithms and the latest improved algorithms, the proposed algorithm achieved an average accuracy of 94.79%, an average precision of 85.77%, an average recall rate of 84.68%, and an F1-score of 84.97% across the 18 datasets. It ranked first in all five evaluation metrics, demonstrating higher stability and accuracy.}, archive = {J_ASOC}, author = {Hui Dong and Caihui Liu and Xiying Chen and Duoqian Miao}, doi = {10.1016/j.asoc.2025.113851}, journal = {Applied Soft Computing}, month = {12}, pages = {113851}, shortjournal = {Appl. Soft. Comput.}, title = {A multi-granularity decision tree algorithm based on variable precision rough sets and zentropy}, volume = {185}, year = {2025}, }
• (2025). Skin lesion classification with mini-batch sampling and deep metric learning. ASOC, 185, 113850. (www)

  @article{ , abstract = {Skin lesion image classification based on deep learning has recently garnered significant attention. However, directly applying methods that perform well in general computer vision tasks to skin lesion image classification is not ideal, as skin lesion image datasets possess intrinsic characteristics, such as class imbalance, intra-class variability, and inter-class similarity. To tackle these challenges simultaneously, we propose a novel unified learning framework, named mBSML, which integrates mini-batch sampling and deep metric learning. In this framework, mini-batch sampling re-samples data in real-time during each iteration of learning, while a new loss function combines mini-batch distance metric-based loss with cross-entropy loss. Through the alternating training procedure on both imbalanced training data and balanced re-sampling data, mBSML effectively learns from global distribution information and local similarity information, not only from the original dataset but also from the minority classes. Extensive experiments conducted on two publicly available datasets demonstrate the effectiveness of mBSML for skin lesion image classification.}, archive = {J_ASOC}, author = {Shengdan Hu and Zhifei Zhang and Li Ying and Guangming Lang}, doi = {10.1016/j.asoc.2025.113850}, journal = {Applied Soft Computing}, month = {12}, pages = {113850}, shortjournal = {Appl. Soft. Comput.}, title = {Skin lesion classification with mini-batch sampling and deep metric learning}, volume = {185}, year = {2025}, }
• (2025). AOD-YOLO: A self-modulating multi-scale feature aggregation mechanism for small object detection in airport surface scenes. ASOC, 185, 113849. (www)

  @article{ , abstract = {Object detection in airport surface scenes is crucial for enhancing safety. However, the coexistence of objects with significant scale disparities within the same region complicates feature representation, limiting existing models’ ability to capture fine-grained details, especially for small objects. To address this challenge, we propose AOD-YOLO, an Airport Object Detection (AOD) model incorporating a Self-Modulating Multi-Scale Feature Aggregation Mechanism. This model introduces two key innovations: (1) Enhanced Context Modeling: By leveraging large-kernel convolution, frequency-domain modulation, and statistical feature analysis, our approach effectively adjusts feature contributions across different object scales, improving contextual understanding in complex scenes; (2) Optimized Small Object Representation: A dynamic gradient gain allocation strategy refines small-object features, enhancing detection accuracy and overall feature presentation. AOD-YOLO consistently improves performance across model scales. On our self-constructed Airport dataset and the public VisDrone-DET2019 dataset, it achieves mean Average Precision (mAP 0.5 ) of 87.9% and 44.9%, respectively—outperforming state-of-the-art models like YOLOv11 and Gold-YOLO by substantial margins. Additionally, through optimized network module placement, AOD-YOLO achieves 112 FPS, striking a balance between computational efficiency and accuracy, making it well-suited for real-time airport object detection.}, archive = {J_ASOC}, author = {Yingqing Wang and Weili Zeng and Ziyu Zhao and Baogeng Li and Zhibin Quan}, doi = {10.1016/j.asoc.2025.113849}, journal = {Applied Soft Computing}, month = {12}, pages = {113849}, shortjournal = {Appl. Soft. Comput.}, title = {AOD-YOLO: A self-modulating multi-scale feature aggregation mechanism for small object detection in airport surface scenes}, volume = {185}, year = {2025}, }
• (2025). Gloss-free sign language translation based on fusion attention. ASOC, 185, 113848. (www)

  @article{ , abstract = {Sign language translation converts sign language videos into spoken language text. Traditional gloss-based approaches require costly gloss annotations, driving recent interest in gloss-free methods. In this paper, we propose a novel gloss-free sign language translation method based on fusion attention (SLTFA) that uniquely models the intrinsic logical structure of sign language. The key innovation is the development of a dual-attention mechanism that mimics the natural hierarchical structure of sign language: intra-gloss attention captures fine-grained relationships within video frame sequences representing individual semantic units, while inter-gloss attention models the broader contextual connections between these units, similar to how words form coherent sentences. Additionally, we introduce a contrastive loss strategy for cross-modal soft alignment that effectively bridges the gap between visual and textual representations. Extensive experiments on the RWTH-PHOENIX-WEATHER-2014T dataset demonstrate SLTFA’s superior performance, achieving a BLEU-4 score of 16.99 and a ROUGE score of 40.82. On the CSL-Daily dataset, our approach achieves a BLEU-1 score of 25.56 and a ROUGE score of 27.51, demonstrating strong performance across different sign languages.}, archive = {J_ASOC}, author = {Yingchun Xie and Wei Su and Chongliang Zhong and Chuan Cai and Yongna Yuan}, doi = {10.1016/j.asoc.2025.113848}, journal = {Applied Soft Computing}, month = {12}, pages = {113848}, shortjournal = {Appl. Soft. Comput.}, title = {Gloss-free sign language translation based on fusion attention}, volume = {185}, year = {2025}, }
• (2025). An efficient mathematical-based optimization method to optimize multi-hydropower operating rules. ASOC, 185, 113846. (www)

  @article{ , abstract = {Optimizing hydropower multi-reservoir systems requires both effective operating rules and efficient optimization techniques. The main contribution of this paper is offering a unique approach that elegantly combines two important parts: creating an efficient optimization method and developing hydropower operating rules. In this regard, a nonlinear rule curve (NLRC) and a linear rule curve (LRC), are tailored for the coordination of a hydropower multi-reservoir system (HMRS) in Iran. To optimize operating rules, the study fabricates a novel algorithm termed the multi-operator weighted mean of vectors (MINFO). The algorithm combines a powerful global search strategy (GSS) that thoroughly searches the solution space with an efficient local search (LS), striking a balance between solution diversity and convergence speed. To fine-tune this balance, an adaptive parameter-tuning strategy is applied. Furthermore, the active-set sequential quadratic programming (ASQP) serves as a localized escaping operator to enhance the algorithm's convergence speed. The effectiveness of the proposed MINFO algorithm is first evaluated through a nonlinear five-reservoir problem. The findings indicate that the MINFO algorithm outperforms a set of 14 distinct optimization methods. Subsequently, the MINFO algorithm is applied to identify optimal NLRC and LRC for a six-reservoir hydropower system. The results underscore the superiority of optimized NLRC, yielding a potential power augmentation of up to 17 % in comparison to the LRC approach. In summation, this study constitutes a seminal contribution by cultivating an efficient rule curve framework for the management of HMRSs.}, archive = {J_ASOC}, author = {Shuguang Li and Iman Ahmadianfar and Aitazaz A. Farooque and Zaher Mundher Yaseen}, doi = {10.1016/j.asoc.2025.113846}, journal = {Applied Soft Computing}, month = {12}, pages = {113846}, shortjournal = {Appl. Soft. Comput.}, title = {An efficient mathematical-based optimization method to optimize multi-hydropower operating rules}, volume = {185}, year = {2025}, }
• (2025). Advanced YOLOv4 for real-time underwater object detection: An application-oriented approach. ASOC, 185, 113837. (www)

  @article{ , abstract = {Detecting debris and monitoring marine life in sea aquaculture face challenges due to limited visibility and the presence of diverse. Underwater object detection by Autonomous Unmanned Vehicle(AUV) is inherently more challenging than land due to light attenuation and water turbidity, especially for small and dense objects in murky images, where extracting high-quality features is hindered. In this paper, we present an efficient approach for real-time underwater object detection through improvements in image enhancement, data augmentation, and feature aggregation. Initially, U-Shape Transformer is applied to enhance the original images. For data augmentation, it is observable that while Mosaic data augmentation enhances complex images but fails to improve small-object detection due generation of less number of images with small objects. To address this limitation, we propose Underwater-Mosaic (U-Mosaic), a modified Mosaic data augmentation technique designed to enhance small-object detection. Additionally, it was noted that existing YOLOv4 struggles with detecting small and densely populated objects in underwater images as unable to get sufficient features for small objects due to downsampling, image quality and also found difficulty in selecting anchor box size. Therefore, we propose a model called Advanced YOLOv4, tailored for underwater object detection. The proposed Advanced YOLOv4 aims to improve object detection efficiency by altering the neck and prediction layers of YOLOv4. Moreover, we introduce an additional spatial pyramid pooling layer to aggregate features and reduce feature dimensions thereby improving object detection rates. Also, the proposed work concentrates on very large object detection and for this purpose used downsampling during the detection of large objects. The proposed approach is validated through two distinct application areas: (i) detecting and locating debris (ii) detecting fish from underwater images. For validation, the Trash ICRA19 dataset is used for debris detection, while the Brackish dataset is employed for fish detection. UIQM and UCIQE, image enhancement assessment metrics are used to measure quality of enhanced images and found more than 20% better result for both the datasets. The proposed real-time underwater object detection model outperformed single-stage object detectors like YOLOv3, YOLOv4, YOLOv5, YOLOv7, and KPE-YOLOv5 by 5% in terms of mean Average Precision(mAP). Also proposed work compared with two-stage detector RCNN and found 8% better mAP than RCNN.}, archive = {J_ASOC}, author = {Pratima Sarkar and Sourav De and Prasenjit Dey and Sandeep Gurung}, doi = {10.1016/j.asoc.2025.113837}, journal = {Applied Soft Computing}, month = {12}, pages = {113837}, shortjournal = {Appl. Soft. Comput.}, title = {Advanced YOLOv4 for real-time underwater object detection: An application-oriented approach}, volume = {185}, year = {2025}, }
• (2025). Multi-branch and multi-loss learning for fine-grained image retrieval. ASOC, 185, 113833. (www)

  @article{ , abstract = {To effectively address the problem of low accuracy of fine-grained image retrieval due to significant intra-class differences and small inter-class differences, we propose a novel and highly reliable fine-grained deep hashing learning framework dubbed MBLNet to accurately retrieve fine-grained images. Specifically, we propose (i) a dual-selected significant region erasure method for generating compact binary codes for fine-grained images; (ii) a dual filtering object location method for mining discriminative local features; and (iii) a new multi-stage loss function for optimizing network training. We conducted extensive experiments on three fine-grained datasets, Stanford Cars, FGVC-Aircraft, and CUB-200-2011, and achieved mAP results of 89.3%, 87.2%, and 80.6%, respectively. Additionally, the ablation study demonstrates that both the dual-selected significant region erasure method and the dual filtering object location method contribute to the improved accuracy of fine-grained image retrieval, further validating the effectiveness of the proposed method. Code can be found at https://github.com/luhongchun/MBLNet.git .}, archive = {J_ASOC}, author = {Hongchun Lu and Min Han and Songlin He and Xue Li and Chase Wu}, doi = {10.1016/j.asoc.2025.113833}, journal = {Applied Soft Computing}, month = {12}, pages = {113833}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-branch and multi-loss learning for fine-grained image retrieval}, volume = {185}, year = {2025}, }
• (2025). UniCon: Unified image-guiding generation with noise consistency. ASOC, 185, 113832. (www)

  @article{ , abstract = {Diffusion models have demonstrated remarkable capabilities in image-to-image tasks. However, existing methods typically focus either on structural (e.g., layout, content) or stylistic guidance, with few approaches effectively excel at both. On the other hand, many methods require time-consuming fine-tuning or high inference latency, making interactive generation applications challenging to realize. To address these issues, we propose a two-stage framework referred as UniCon ( Uni fied Image-guiding Generation with Noise Con sistency). To improve time efficiency, we follow the paradigm of inversion-based image manipulation and introduce a novel method called Noise Consistency Inversion . Leveraging the nature of Consistency Models, this inversion process is highly efficient, requiring only a single neural function evaluation (NFE) in the inversion process. To achieve high consistency and finer control, we introduce a unified attention-based guidance mechanism that supports structural, stylistic, or joint reference inputs, without any additional fine-tuning. Experiments with structure- and style-specific methods show that our approach performs competitively or better in each individual aspect. In comparison of style transfer tasks that demand both structure and style, our method outperforms state-of-the-art baselines, confirming the effectiveness of our union control strategy. And overall, our approach also achieves the best efficiency in terms of runtime performance.}, archive = {J_ASOC}, author = {Yuanjun Liao and Yuning Gong and Yanci Zhang}, doi = {10.1016/j.asoc.2025.113832}, journal = {Applied Soft Computing}, month = {12}, pages = {113832}, shortjournal = {Appl. Soft. Comput.}, title = {UniCon: Unified image-guiding generation with noise consistency}, volume = {185}, year = {2025}, }
• (2025). Evaluation of the financial performance of the textile and apparel industry in interval type-2 fuzzy environment. ASOC, 185, 113830. (www)

  @article{ , abstract = {The Turkish textile and apparel sector plays a crucial role in the national economy through employment, exports, and investment. The financial performance of companies is a key determinant of their sustainability and competitiveness, especially in global markets. The Turkish textile and apparel sector is one of the essential industries in terms of macro-economic indicators such as net foreign exchange inflow, employment and investment. This sector is also one of the critical actors in world trade. A robust performance evaluation model is essential for stakeholders such as investors, creditors, and managers. However, the assessment of firms is a very critical decision involving uncertainty due to various conflicting criteria based on judgements. In this study, an integrated multi-criteria decision-making (MCDM) model including interval type-2 fuzzy hierarchy process (IT2FAHP) and Compromise Ranking of Alternatives from Distance to Ideal Solution (CRADIS) approaches are proposed to assess the financial performance of Turkish textile and clothing firms that are traded in Borsa İstanbul (BİST) in the period from 2006 to 2020. In line with the determined purpose, the arithmetic average of the determined financial ratios during the analysis period covering 15 years is computed to obtain long-term performance indicators. The importance weights of the selected financial criteria for the performance evaluation model are identified by employing the IT2FAHP approach. Then, the firms are ranked according to their financial performances with the CRADIS method. In addition, the results from the sensitivity analysis validate the proposed approach and prove that it is practical. Moreover, practical and managerial implications are discussed based on the results. The results offer valuable insights for strategic decision-making and can support efforts to enhance financial stability in the textile and apparel sector. According to the results, "LUKSK" had the highest long-term financial performance among the 11 companies discussed. This company is followed by BOSSA, YATAS, and ATEKS companies. The alternatives confirm the robustness of the proposed model in maintaining its place in the ranking in 190 scenarios. In addition, the comparative analysis confirms the consistency of the proposed ranking framework.}, archive = {J_ASOC}, author = {Ömer Faruk Görçün and Mohsin Shabir and Ahmet Çalık and Özcan Işık}, doi = {10.1016/j.asoc.2025.113830}, journal = {Applied Soft Computing}, month = {12}, pages = {113830}, shortjournal = {Appl. Soft. Comput.}, title = {Evaluation of the financial performance of the textile and apparel industry in interval type-2 fuzzy environment}, volume = {185}, year = {2025}, }
• (2025). An interpretable dual-output multivariate wind speed interval prediction scheme using hyper-heuristic optimization algorithm and deep neural networks. ASOC, 185, 113829. (www)

  @article{ , abstract = {Uncertainty analysis of wind speed forecasting using the Lower Upper Bound Estimation (LUBE) represents an advanced interval prediction method that does not require assumptions about data distribution. Previous studies, however, have exclusively focused on univariate prediction models, neglecting the information from other variables, and have not fully exploited the prediction errors in their loss function during training. To address these issues, an interpretable dual-output multivariate wind speed interval prediction scheme (IMWSIPS) that utilizes a hyper-heuristic optimization algorithm and a deep neural network is proposed, along with a novel loss function for training. The system initially takes multiple inputs such as historical wind speed and other influencing factors including wind direction, density, temperature, and pressure into a deep neural network. The actual wind speeds are then scaled up and down by factors of 1 + θ 1 (0 <θ 1 <1) and 1 + θ 2 (-1 <θ 2 <0), respectively, to produce two outputs from the network. On this basis, an optimization problem to minimize interval width under a given coverage probability is formulated and solved using the developed hyper-heuristic algorithm, yielding optimal values for θ 1 and θ 2 and the prediction intervals for sub-models. Subsequently, the advantages of five deep neural network models are leveraged to construct an ensemble model, with weights optimized by the hyper-heuristic algorithm to derive the final prediction intervals. Ultimately, the system's interpretability is analyzed at both variable and sub-model levels. Experimental and discussion results demonstrate that the introduction of IMWSIPS not only signifies enhancements in forecasting performance but also implies improvements in wind energy utilization efficiency and reductions in operational costs for power systems.}, archive = {J_ASOC}, author = {Mengzheng Lv and Jianzhou Wang and Shuai Wang and Yang Zhao and Jialu Gao and Yuansheng Qian}, doi = {10.1016/j.asoc.2025.113829}, journal = {Applied Soft Computing}, month = {12}, pages = {113829}, shortjournal = {Appl. Soft. Comput.}, title = {An interpretable dual-output multivariate wind speed interval prediction scheme using hyper-heuristic optimization algorithm and deep neural networks}, volume = {185}, year = {2025}, }
• (2025). The fuzzified grey wolf: An improved grey wolf optimizer based on dynamic fuzzy system FGWO. ASOC, 185, 113818. (www)

  @article{ , abstract = {The Grey Wolf Optimizer (GWO) is a leading, powerful, and effective algorithm in swarm optimization techniques, showing competitive performance across various optimization problems. Yet, GWO is burdened by high tendencies toward exploitation and imprecise population diversity. This work introduces an improved GWO called the Fuzzified Grey Wolf Optimizer (FGWO) for solving global optimization problems. FGWO benefits from a dynamic fuzzy inference system (DFIS) to capture the optimal value of a → throughout iterations. DFIS integrates two inputs: the diversity rate and iteration number, and by inferring the optimal value of a → , FGOW determines whether to exploit or explore. Moreover, DFIS employs an adaptive membership function to capture the precise value of population diversity throughout the course of iteration. This optimal a → determination strategy can achieve a balanced exploitation–exploration ratio, mitigating premature convergence and enhancing diversity. FGWO is evaluated on CEC2017 benchmark functions, four engineering designs, and a breast cancer genes feature selection design. FGWO was compared with three other improved GWOs, and across all experiments, the outcomes demonstrate its superiority in terms of efficiency and applicability to real-world designed problems.}, archive = {J_ASOC}, author = {Mohammed Dheyaa Algubili and Labiba M. Alhelfi and Hana’ M. Ali}, doi = {10.1016/j.asoc.2025.113818}, journal = {Applied Soft Computing}, month = {12}, pages = {113818}, shortjournal = {Appl. Soft. Comput.}, title = {The fuzzified grey wolf: An improved grey wolf optimizer based on dynamic fuzzy system FGWO}, volume = {185}, year = {2025}, }
• (2025). Fuzzy lagrange interpolation method from summation of interactive fuzzy numbers. ASOC, 185, 113817. (www)

  @article{ , abstract = {This article proposes a novel approach to extending the sum of interactive fuzzy numbers, which is independent of the order of its operands. Interactive fuzzy numbers are fuzzy quantities in which the values across different α -cuts are not assumed to vary independently, incorporating dependencies that better reflect real-world uncertainty. A characterization of the proposed summation is given in terms of α -cuts, making computational implementation easier. It is shown that this operation preserves essential mathematical properties, including associativity. This is particularly important, as it enables the consistent aggregation of multiple fuzzy quantities without concern for the order in which the operands are grouped. The norm and width behaviors under this new summation are also analyzed. To illustrate the theoretical results, several examples are provided. As a practical application, the classical Lagrange polynomial interpolation method is extended to handle uncertain parameters represented by interactive fuzzy numbers. A fuzzy curve fitting problem is examined using this framework, and a comparative discussion highlights the advantages of the proposed method over existing approaches.}, archive = {J_ASOC}, author = {Geizane Lima da Silva and Estevão Esmi and Vinícius Francisco Wasques and Laécio Carvalho de Barros}, doi = {10.1016/j.asoc.2025.113817}, journal = {Applied Soft Computing}, month = {12}, pages = {113817}, shortjournal = {Appl. Soft. Comput.}, title = {Fuzzy lagrange interpolation method from summation of interactive fuzzy numbers}, volume = {185}, year = {2025}, }
• (2025). Decision analytics for indian culinary tourism: A holistic group approach considering correlation. ASOC, 185, 113812. (www)

  @article{ , abstract = {In today’s data-driven world, making informed decisions in dynamic fields like culinary tourism is crucial. An enhanced multi-attribute decision-making (MADM) model is presented in this study to tackle the uncertainty and interdependencies of India’s culinary tourism landscape. The main goals are to (i) address uncertainty and correlations in MADM scenarios, (ii) calculate objective attribute weights, and (iii) resolve conflicts among alternatives based on preference, indifference, and incomparability. To manage uncertainty, the proposed model incorporates r , s -quasirung orthopair fuzzy set ( r , s -QOFS), while to capture relational dynamics among factors Aczel–Alsina operations based geometric Heronian mean operator is developed. Attribute weighting is performed with MEREC (method based on the removal effects of criteria) method, while a modified ORESTE (organísation, rangement et Synthèse dedonnées relarionnelles (in French)) method within the r , s -QOFS is initiated to rank alternatives, introducing a new ranking measure in place of Besson’s traditional rank. Finally, to test the effectiveness and practical value, a case study of culinary tourism destinations across 36 Indian states and union territories is conducted and then ranked using the proposed model. The results highlight southern Indian states as preferred destinations. Thus, this work contributes in two ways: first, by providing a general decision-making model for imprecise and data, and second, by offering valuable insights into the future of Indian culinary tourism.}, archive = {J_ASOC}, author = {Kaushik Debnath and Sankar Kumar Roy}, doi = {10.1016/j.asoc.2025.113812}, journal = {Applied Soft Computing}, month = {12}, pages = {113812}, shortjournal = {Appl. Soft. Comput.}, title = {Decision analytics for indian culinary tourism: A holistic group approach considering correlation}, volume = {185}, year = {2025}, }
• (2025). Bridging the semantic-numerical gap: A numerical reasoning method of cross-modal knowledge graph for material property prediction. ASOC, 185, 113776. (www)

  @article{ , abstract = {Material properties are illustrated by numerical data and semantic factors. In general, existing methods typically adopt machine learning (ML) algorithms to regress numerical properties or transfer other pre-trained knowledge graphs (KGs) to the material, due to the limitations of small-sample datasets. However, integrating semantic and numerical information from multi-modal data which across diverse experimental conditions remains a significant challenge in materials science. In this paper, a numerical reasoning method for material KGs (NR-KG) 1 was proposed, which constructs a cross-modal KG using semantic nodes and numerical proxy nodes. Both types of information by projecting KG into a canonical KG were captured and a graph neural network to predict material properties was utilized. In process, a novel projection prediction loss is proposed to extract semantic features from numerical information. NR-KG facilitates end-to-end processing of cross-modal data, mining relationships and cross-modal information in small-sample datasets, and fully utilizes effective experimental data to enhance the accuracy of material prediction. We propose two new high-entropy alloys (HEA) property datasets with semantic descriptions. NR-KG outperforms state-of-the-art (SOTA) methods on two material datasets, with MSE values of 3520 and 2.210, and achieving relative improvements of 25.9% and 16.1%, respectively, over the second-best methods, KANO and PCHMLP (semantic). It also achieves RMSE values of 0.584 and 0.521 on the FreeSolv and ESOL public molecular datasets, surpassing SOTA methods by 48.8% and 22.2% over KANO, highlighting its potential application and generalizability.}, archive = {J_ASOC}, author = {Guangxuan Song and Dongmei Fu and Zhongwei Qiu and Zijiang Yang and Jiaxin Dai and Lingwei Ma and Dawei Zhang}, doi = {10.1016/j.asoc.2025.113776}, journal = {Applied Soft Computing}, month = {12}, pages = {113776}, shortjournal = {Appl. Soft. Comput.}, title = {Bridging the semantic-numerical gap: A numerical reasoning method of cross-modal knowledge graph for material property prediction}, volume = {185}, year = {2025}, }
• (2025). Transformer-based actor–critic for adaptive cryptocurrency portfolio rebalancing. ASOC, 185, 113697. (www)

  @article{ , abstract = {High-frequency trading in volatile markets, such as cryptocurrencies, requires portfolio models that can swiftly adapt to regime shifts while controlling risk. We propose a novel approach that frames portfolio management as a dynamic strategy-selection problem. Instead of directly predicting asset weights, our agent selects from a pool of expert strategies based on recent market trends. We introduce a Transformer-based Variational Autoencoder (VAE) to extract disentangled trend representations, and a trend-aware actor–critic model to perform expert selection. Experiments demonstrate that this modular, strategy-level control mechanism outperforms existing methods in risk-sensitive crypto portfolio management.}, archive = {J_ASOC}, author = {Ahmad Asadi and Reza Safabakhsh}, doi = {10.1016/j.asoc.2025.113697}, journal = {Applied Soft Computing}, month = {12}, pages = {113697}, shortjournal = {Appl. Soft. Comput.}, title = {Transformer-based actor–critic for adaptive cryptocurrency portfolio rebalancing}, volume = {185}, year = {2025}, }
• (2025). Sustainable bridge management using refined feature selection for machine learning-aided bridge condition prediction: Incorporation of pareto distribution in MRMR method. ASOC, 184, 113878. (www)

  @article{ , abstract = {One of the critical phases of bridge management is condition prediction, which was enhanced after machine learning emerged. Researchers have used feature selection (FS) methods to reduce the predictors and optimise the prediction models. Prior studies have either explored a unified feature set for the overall bridge condition or focused solely on the deck, leading to limited predictions. This study proposes a refined feature selection method highlighting the importance of specific predictors for different bridge elements’ conditions using the California State inspection database from 1983 to 2021. The implemented FS approach consists of a verified Minimum-Redundancy Maximum-Relevance (MRMR) method followed by a Pareto analysis that identifies the most contributory factors in predicting the condition of bridges’ decks, superstructures, and substructures. The applied experiment on the US National Bridge Inventory database reveals that 28–33 predictor variables, out of more than 140 available features, contribute the most to each component’s health prediction with a cumulative importance score of over 95 %. Additionally, 22 mutual data items among the selected features are proposed as the minimum required predictors to be gathered by the asset management authorities. This study’s achievements help both researchers reduce the running costs of their prediction models and asset managers with data gathering and registration optimisation and, consequently, whole-of-life cycle cost reduction for sustainable asset management.}, archive = {J_ASOC}, author = {Vandad Dayan and Nicholas Chileshe and Reza Hassanli and Amin Parvaneh}, doi = {10.1016/j.asoc.2025.113878}, journal = {Applied Soft Computing}, month = {12}, pages = {113878}, shortjournal = {Appl. Soft. Comput.}, title = {Sustainable bridge management using refined feature selection for machine learning-aided bridge condition prediction: Incorporation of pareto distribution in MRMR method}, volume = {184}, year = {2025}, }
• (2025). Development of fuzzy rule-based models in the presence of the big data environment. ASOC, 184, 113869. (www)

  @article{ , abstract = {In this study, we propose a series of methods to build fuzzy rule-based models (FRBMs) in the presence of the big data environment such that the formed predictive models are more accurate, efficient, and robust. We follow two major steps to realize this target. In the first step, we build numeric FRBMs with the big data set such that the formed predictive models are more accurate and efficient. Specifically, based on the divide-and-conquer strategy, the big data set is divided into subsets through either the hyperplane division-based method or the K -Means clustering-based method; then either a global-based strategy or a local-based strategy is used to build numeric FRBMs. As a result, four Options are generated to develop numeric FRBMs. In the second step, we build the granular FRBMs based on the four Options developing the numeric FRBMs. Specifically, given a certain Option, based on the Principle of Justifiable Granularity (PJG), we granulate both condition parts and conclusion parts of the rules, forming the granular FRBMs; then the predictive models are further evaluated based on the PJG and optimized based on the Particle Swarm Optimization (PSO) algorithm to enhance the robustness. Finally, experimental studies on both synthetic datasets and publicly available datasets are conducted to prove the effectiveness of the proposed methods.}, archive = {J_ASOC}, author = {Yinghua Shen and Dan Zhao and Yan Li and Xingchen Hu and Yuan Chen and Bingsheng Liu}, doi = {10.1016/j.asoc.2025.113869}, journal = {Applied Soft Computing}, month = {12}, pages = {113869}, shortjournal = {Appl. Soft. Comput.}, title = {Development of fuzzy rule-based models in the presence of the big data environment}, volume = {184}, year = {2025}, }
• (2025). CMFF: A cross-modal multi-layer feature fusion network for multimodal sentiment analysis. ASOC, 184, 113868. (www)

  @article{ , abstract = {Multimodal sentiment analysis seeks to interpret speaker sentiment by integrating information from multiple modalities, typically text and audio. While existing methods often focus on fusing deep-layer features extracted from the final stages of unimodal encoders, they may overlook crucial fine-grained information present in shallow-layer features (e.g., subtle phonetic variations or basic syntactic structures) relevant for nuanced sentiment understanding. Furthermore, effectively fusing features from different modalities presents the dual challenges of dynamically weighting each modality’s contribution and accommodating their inherent data heterogeneity. To address these limitations, we propose a novel Cross-modal Multi-layer Feature Fusion (CMFF) network. CMFF explicitly leverages the hierarchical information contained in both shallow-layer and deep-layer features from text and audio modalities. It employs multi-head cross-modal attention mechanisms within its fusion layers to facilitate interaction across feature layers and modalities. Crucially, CMFF incorporates a Mixture of Gated Experts (MoGE) network within these fusion layers. The MoGE utilizes modality-specific expert sub-networks, each tailored to process the distinct characteristics of text or audio data, thereby directly addressing data heterogeneity. Concurrently, each expert employs an internal gated feed-forward mechanism. This allows the model to dynamically control the information flow for each feature vector, effectively learning to weigh the importance of different feature dimensions from each layer and modality based on the input context. Extensive experiments conducted on the benchmark CMU-MOSI and CMU-MOSEI datasets demonstrate that the proposed CMFF model achieves competitive or superior performance compared to state-of-the-art methods across various standard evaluation metrics.}, archive = {J_ASOC}, author = {Shuting Zheng and Jingling Zhang and Yuanzhao Deng and Lanxiang Chen}, doi = {10.1016/j.asoc.2025.113868}, journal = {Applied Soft Computing}, month = {12}, pages = {113868}, shortjournal = {Appl. Soft. Comput.}, title = {CMFF: A cross-modal multi-layer feature fusion network for multimodal sentiment analysis}, volume = {184}, year = {2025}, }
• (2025). Mitigating over-squashing in graph few-shot learning by leveraging local and global similarities. ASOC, 184, 113863. (www)

  @article{ , abstract = {Supervised machine learning models, particularly neural networks, often fail to deliver satisfactory results in scenarios with insufficient data. This becomes even more challenging when dealing with inherently complex data, such as graph data. This paper addresses the issue of learning with a limited number of samples, known as n -way k -shot learning, within the context of graph data. Our research extends the concept of similarity from neighboring nodes to the entire graph by leveraging transitivity relations. By employing edges and strong transitivity relations, we utilize a bipartite graph neural network that capitalizes on both local neighborhoods and distant, yet similar, nodes to generate node embeddings. This approach has demonstrated effectiveness in tasks such as node classification. Our proposed model’s ability to mitigate the over-squashing problem enhances its generalizability, resulting in a task-invariant model. Experimental results on various graph datasets show that the embeddings produced by our model are not task-specific. Consequently, our model outperforms other models in few-shot learning scenarios, where only a limited number of labeled nodes are available for each distinct downstream task.}, archive = {J_ASOC}, author = {Yassin Mohamadi and Mostafa Haghir Chehreghani}, doi = {10.1016/j.asoc.2025.113863}, journal = {Applied Soft Computing}, month = {12}, pages = {113863}, shortjournal = {Appl. Soft. Comput.}, title = {Mitigating over-squashing in graph few-shot learning by leveraging local and global similarities}, volume = {184}, year = {2025}, }
• (2025). Causal intervention for feedback information in fairness recommendation. ASOC, 184, 113862. (www)

  @article{ , abstract = {In recommender systems, users’ feedback information is greatly affected by individual differences between users and usually interferes with the interaction between users and items, causing unfairness in the recommendation results. To alleviate the problem of unfair recommendation results caused by the confounding of feedback information, we propose a novel causal intervention for feedback information in fairness recommendation (CIFair). First, we construct a causal graph based on the interaction between users and items and analyze the reasons for unfair recommendation results caused by feedback information as the confounding factor through the causal graph. Then, we design a two-phase predicted rating generation, namely, the elimination and reconstruction phases. In the elimination phase, we analyze the dual effects of the confounding factor on the recommender system and eliminate it through causal intervention to improve the fairness of the recommendation results and obtain a fair predicted rating. In the reconstruction phase, we design personalized feedback information based on user and item attributes to ensure recommendation performance and obtain a personalized predicted rating. Finally, we combine the predicted ratings generated in the elimination and reconstruction phases and provide users with personalized recommendation results with fairness. We conduct experiments on three publicly available datasets (MovieLens-1M, Last.fm, and Yelp) to verify the significance of the CIFair. The ablation experiment confirms that the CIFair model can achieve relatively fair recommendations from both the user and item sides while ensuring recommendation performance.}, archive = {J_ASOC}, author = {Chenyu Wang and Guanxi Wang and Guowei Yang and Dun Li}, doi = {10.1016/j.asoc.2025.113862}, journal = {Applied Soft Computing}, month = {12}, pages = {113862}, shortjournal = {Appl. Soft. Comput.}, title = {Causal intervention for feedback information in fairness recommendation}, volume = {184}, year = {2025}, }
• (2025). Fixing deep early exit ensembles for sensor-based human activity recognition through uncertainty quantification. ASOC, 184, 113861. (www)

  @article{ , abstract = {Dynamic deep neural networks (DNNs) have recently achieved outstanding success on a wide range of resource-constrained human activity recognition (HAR) tasks. However, prior most works are deterministic and could not provide any uncertainty estimate. Though an early exit can facilitate adaptive activity inference by producing intermediate predictions at multiple stages during forward pass, these predictions are only meaningful in real-world while complemented with reliable uncertainty estimates. Until now, the quality of uncertainty estimates has always been ignored in the context of early exit HAR. How to quantify predictive uncertainty in dynamic DNNs still remains challenging and yet unsolved. To address this issue, this paper introduces a new framework of early exit ensembles, which provides a probabilistic treatment of such dynamic DNNs to capture uncertainty estimates through an implicit ensemble of sub-networks sharing weights. We evaluate the proposed approach using three strong state-of-the-art DNN backbones on several mainstream HAR benchmarks, i.e., UCI-HAR, UniMib-SHAR and WISDM. Depending on the backbones and datasets, our approach can lower calibration error up to around 11 × , while increasing accuracy by up to 0.588% over its single counterpart. Both theoretical computational efficiency and practical runtime latency are analyzed. We provide an intuitive illustration of accounting for both aleatoric and epistemic uncertainty, which validates that such probabilistic treatment can adequately capture uncertainty estimates to aid decision-making while varying the computational budgets.}, archive = {J_ASOC}, author = {Xin Liu and Lei Zhang and Wenbo Huang and Dongzhou Cheng and Hao Wu and Aiguo Song}, doi = {10.1016/j.asoc.2025.113861}, journal = {Applied Soft Computing}, month = {12}, pages = {113861}, shortjournal = {Appl. Soft. Comput.}, title = {Fixing deep early exit ensembles for sensor-based human activity recognition through uncertainty quantification}, volume = {184}, year = {2025}, }
• (2025). A transformer-based water quality prediction model with mixed fractional brownian features and multi-feature bottleneck transformation. ASOC, 184, 113860. (www)

  @article{ , abstract = {Data-driven water-quality model inference typically depends on large datasets. In emergency or resource-constrained settings, however, small sample sizes hinder a model’s ability to faithfully capture the complex nonlinear and multiscale dynamics of water quality. To address this challenge, we propose an improved Transformer. First, we enrich the inputs with features from mixed fractional Brownian motion with a perturbation factor (rMFBM). The rMFBM module captures heterogeneous temporal dependencies via multiple Hurst exponents and applies a Cholesky-based covariance sampler with a diagonal perturbation to ensure numerical stability. Second, a multi-feature bottleneck layer performs compression and dimensionality reduction to yield compact yet information-dense representations. Finally, the optimized features are modeled by a lightweight Transformer trained with the AMSGrad optimizer and an early-stopping strategy. On the task of predicting the N S F W Q I water-quality index, the proposed model reduces root-mean-square error (RMSE) by approximately 43.1% relative to a standard Transformer and improves the coefficient of determination ( R 2 ) by 1.89% across four small-sample datasets. It also outperforms widely used alternatives, including GNNs and PINNs. These results suggest that certain water-quality parameters exhibit mixed fractional Brownian-motion characteristics, enabling rapid and reliable prediction under data scarcity and in emergency scenarios.}, archive = {J_ASOC}, author = {Genghao Cui and Zhiyao Zhao and Li Wang and Huiyan Zhang and Jiabin Yu}, doi = {10.1016/j.asoc.2025.113860}, journal = {Applied Soft Computing}, month = {12}, pages = {113860}, shortjournal = {Appl. Soft. Comput.}, title = {A transformer-based water quality prediction model with mixed fractional brownian features and multi-feature bottleneck transformation}, volume = {184}, year = {2025}, }
• (2025). MPCMamba: Privacy-preserving inference for mamba models via secure multi-party computation. ASOC, 184, 113859. (www)

  @article{ , abstract = {Deep learning is increasingly critical across high-stakes domains, but privacy risks during model inference remain a major concern. Secure multi-party computation (SMPC) offers a promising solution by enabling privacy-preserving collaborative inference without exposing sensitive data. The recently proposed Mamba model, which outperforms Transformer in certain tasks, presents unique challenges for SMPC due to its state-space architecture and nonlinear operations. This paper introduces a framework for executing Mamba model inference under SMPC while preserving privacy. The framework natively supports linear operations and securely computes nonlinear functions – including square roots, exponentials, logarithms, and SiLU activations – without altering the original model architecture. Experimental results demonstrate that the proposed method achieves accuracy improvements of 2.4%, 2.84%, and 8.23% on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets, respectively, compared to existing SMPC-based vision Transformer approaches, MPCViT. Additionally, inference latency is reduced by factors of 2.14 × , 2.14 × , and 26.13 × on these benchmarks, significantly advancing efficient and secure deployment of state-space models in privacy-sensitive scenarios.}, archive = {J_ASOC}, author = {Yongqiang Yu and Yuliang Lu and Xuehu Yan and Wei Yan and Shengyang Luo}, doi = {10.1016/j.asoc.2025.113859}, journal = {Applied Soft Computing}, month = {12}, pages = {113859}, shortjournal = {Appl. Soft. Comput.}, title = {MPCMamba: Privacy-preserving inference for mamba models via secure multi-party computation}, volume = {184}, year = {2025}, }
• (2025). A depth-estimation-based method for multi-view synthesis applied to chinese landscape paintings. ASOC, 184, 113858. (www)

  @article{ , abstract = {This paper addresses the challenging task of novel view synthesis for traditional Chinese landscape paintings, which typically offer only a single perspective and lack clear depth information. To overcome the limitations of existing methods that rely on multi-view input and depth estimation, we propose a multi-view synthesis method for Chinese landscape paintings, termed MVSM-CLP. The proposed CLPDepth Module employs a high-low resolution fusion mechanism to enhance detail expression while preserving the original scene structure. We introduce an image restoration technique guided by landscape ink lines, termed LInpainting, to improve edge extraction and the accuracy of painting inpainting Additionally, our method tackles the issue of scarce 3D data in current view synthesis efforts by constructing multi-view data from a single ancient painting. Our approach effectively bridges the gap between 2D art and 3D visualization, creating vivid and realistic virtual environments while preserving the traditional style and essence of Chinese paintings. Experimental results demonstrate the effectiveness of our method in achieving high-quality multi-view synthesis, offering new possibilities for the digital preservation of cultural heritage. A preprint has previously been published (Peng et al., 2024 [1] ). The code and dataset is available at https://github.com/LPDLG/DepCLP_Dataset .}, archive = {J_ASOC}, author = {Xianlin Peng and Wanlin Zhou and Qiyao Hu and Tengfei Li and Dong Zhang and Rui Cao}, doi = {10.1016/j.asoc.2025.113858}, journal = {Applied Soft Computing}, month = {12}, pages = {113858}, shortjournal = {Appl. Soft. Comput.}, title = {A depth-estimation-based method for multi-view synthesis applied to chinese landscape paintings}, volume = {184}, year = {2025}, }
• (2025). SGBA: Subspace guidance backdoor attack with feature alignment in image classification. ASOC, 184, 113857. (www)

  @article{ , abstract = {With the widely use of deep neural networks (DNNs) in a variety of classification and generation tasks with remarkable performance, the security of DNN models has attracted further attention. Recently, backdoor attacks have made a significant threat to DNN models, where attackers intentionally introduce malicious patterns into the model and manipulate the behavior of the model upon encountering the specific trigger. In this paper, we analyze the workflow of backdoor attacks with feature-guide classifier, and propose a novel method to conduct the backdoor attack based on feature alignment. We find that using samples with adversarial perturbation for training can mislead the model during the inference stage. Inspired by this observation, we leverage the intrinsic distribution of target class features in the latent space to generate effective and invisible triggers by adding directional perturbations to target images. We evaluate the attack under three widely used datasets and results show that our method can achieve considerable performance in comparison with other four state-of-the-art attacks. Furthermore, we also make extensive experiment to emphasize the robustness and stealthiness of our attack method.}, archive = {J_ASOC}, author = {Hao Luo and Zhi Qin and Lin Wang and Ziyue Wu and Min Yang}, doi = {10.1016/j.asoc.2025.113857}, journal = {Applied Soft Computing}, month = {12}, pages = {113857}, shortjournal = {Appl. Soft. Comput.}, title = {SGBA: Subspace guidance backdoor attack with feature alignment in image classification}, volume = {184}, year = {2025}, }
• (2025). Deep learning-based kiwifruit flower recognition method to facilitate automated pollination. ASOC, 184, 113855. (www)

  @article{ , abstract = {Accurate flowering-stage recognition is vital for intelligent orchard management, automated pollination, and yield forecasting. However, in complex natural scenes, existing models struggle to balance detection accuracy with inference speed. To bridge this gap, we propose the Kiwifruit Recognition Network (KiwiRecNet), a lightweight yet high-performance framework tailored to kiwifruit blossoms in the wild. KiwiRecNet first employs the Kiwifruit Generative Adversarial Network for Low-light Improvement (KiwiGAN-LI) to enhance under-exposed images. We then design a novel backbone, the Multi-Scale Shuffle Block (MSBlock), which combines structural re-parameterisation with channel–spatial shuffling to shrink the network footprint. Next, we propose the Partial-Mixing Vision Transformer (PMVIT), a convolution-Transformer hybrid that captures fine-grained features and remains robust to occlusion. Finally, we devise a Bidirectional Cross-Scale Fusion module (Bi-CSF) to enrich multiscale perception. Evaluated on the NWAFU Kiwifruit_F dataset, KiwiRecNet achieves 94.07 % mAP at 82.67 FPS with only 0.93 million parameters, outperforming existing lightweight detectors while approaching heavyweight baselines at a fraction of their cost. Consistent gains across multiple flower datasets confirm its generalisation ability. These results demonstrate an effective route to high-accuracy, real-time flowering-phase recognition on resource-constrained devices, paving the way for scalable agricultural automation.}, archive = {J_ASOC}, author = {Xiaopeng Li and Jinzhi Du and Xiaoyu Chen and Fuxi Shi and Shuqin Li}, doi = {10.1016/j.asoc.2025.113855}, journal = {Applied Soft Computing}, month = {12}, pages = {113855}, shortjournal = {Appl. Soft. Comput.}, title = {Deep learning-based kiwifruit flower recognition method to facilitate automated pollination}, volume = {184}, year = {2025}, }
• (2025). Searching for local pareto fronts based on the non-dominance range in the decision space. ASOC, 184, 113853. (www)

  @article{ , abstract = {Multi-modal multi-objective optimisation problems (MMOPs) involve multiple equivalent Pareto sets that share the same Pareto front, and a special subclass is known as MMOPs with local Pareto fronts (MMOPLs). Conventional multi-modal multi-objective optimisation evolutionary algorithms (MMOEAs) struggle to effectively identify local Pareto fronts, and existing methods tailored for MMOPLs exhibit notable limitations. Additionally, commonly used performance metrics lack systematic evaluation and suffer from inherent shortcomings. A simple yet effective MMOEA that leverages the non-dominance range in the decision space is proposed to address these challenges. By prioritising individuals with above-average non-dominance ranges, the algorithm enhances the identification and retention of both global and local optima. Convergence is further improved using the local outlier factor method. The limitations of existing performance metrics are analysed and six new performance metrics tailored for MMOPLs are introduced. To facilitate evaluation, benchmark problems are modified to create scenarios in which global and local optima coexist on the same Pareto set. Extensive experimental results confirm that the proposed algorithm achieves competitive performance, effectively identifying global and local optima while ensuring well-distributed solutions.}, archive = {J_ASOC}, author = {Yimin Shen and Yu Guo and Shaohua Huang and Weiwei Qian and Shengbo Wang and Litong Zhang}, doi = {10.1016/j.asoc.2025.113853}, journal = {Applied Soft Computing}, month = {12}, pages = {113853}, shortjournal = {Appl. Soft. Comput.}, title = {Searching for local pareto fronts based on the non-dominance range in the decision space}, volume = {184}, year = {2025}, }
• (2025). Normalizing flow defect detection model based on similar self-supervision. ASOC, 184, 113847. (www)

  @article{ , abstract = {To overcome the limitations in unsupervised industrial anomaly detection, such as poor modeling of the appearance differences between synthetic and real defects and the failure to capture multi-scale feature relationships, this paper proposes a Normalizing Flow Defect Detection Model Based on Similar Self-Supervision (NF-SS). First, a feature-level defect generation method is introduced. It synthesizes realistic defect samples by fusing defect background features, known defect features from related domains, and other auxiliary features. This enhances the model's understanding of defect characteristics and reduces the gap between synthetic and real defects. Second, a multi-scale joint Normalizing Flow decoder is proposed. It replaces independent NF layers with progressive multi-scale fusion, allowing the model to integrate features hierarchically across scales. This preserves spatial relationships, reduces edge blurring, and improves defect localization accuracy. Extensive experiments on the MVTec AD dataset show that NF-SS outperforms state-of-the-art models, achieving an average image-level AUC of 99.54 % and pixel-level AUC of 98.76 %. It significantly improves the detection of subtle and ambiguous defects. NF-SS combines unsupervised and self-supervised learning, providing a robust solution for industrial quality inspection with limited defect samples.}, archive = {J_ASOC}, author = {Zhenlian Miao and Guangzhu Chen and Herui Cao and Yuan Tang and Xiaojuan Liao}, doi = {10.1016/j.asoc.2025.113847}, journal = {Applied Soft Computing}, month = {12}, pages = {113847}, shortjournal = {Appl. Soft. Comput.}, title = {Normalizing flow defect detection model based on similar self-supervision}, volume = {184}, year = {2025}, }
• (2025). Refined vision to obtain the vibration trajectory of rotating body. ASOC, 184, 113845. (www)

  @article{ , abstract = {Non-contact vision technology has been widely used in many fields, but it still has limited accuracy in dealing with complex motion, small displacement and boundary recognition, especially in the rotating body scene, where it is difficult to balance global robustness and local sensitivity. To achieve high-precision visual measurement of rotating structure vibration, this paper proposes a detection-guided target tracking method to address the shortcomings of existing visual algorithms in target identity preservation and temporal consistency. In the face of the above challenges, this paper constructs lightweight encoder–decoder network to extract multi-scale semantic information to enhance the target edge modeling capability, and designs a multidimensional regression mechanism to predict the target boundary from the center point and extract sub-pixel vibration signals. At the same time, a guide term is introduced to optimize trajectory association and improve the temporal continuity and spatial consistency of tracking. Experimental results show that the average performance of the proposed method is improved by about 51.2% on the basis of multiple mainstream visual algorithms, and the main frequency recognition error is reduced to 0.32 Hz, which effectively suppresses high-frequency vibration errors. Ablation experiments verify the contribution of each module to the measurement accuracy, and low-quality image tests further show that the method has good generalization ability and application prospects.}, archive = {J_ASOC}, author = {Rongliang Yang and Tao Liu and Sen Wang}, doi = {10.1016/j.asoc.2025.113845}, journal = {Applied Soft Computing}, month = {12}, pages = {113845}, shortjournal = {Appl. Soft. Comput.}, title = {Refined vision to obtain the vibration trajectory of rotating body}, volume = {184}, year = {2025}, }
• (2025). Improving lightweight semi-supervised text classification via teacher intervention. ASOC, 184, 113844. (www)

  @article{ , abstract = {The most recent semi-supervised text classification frameworks have focused on pre-trained models like BERT. While effective, their large-scale parameters and slow inference speed hinder deployment in many practical scenarios. In this work, we develop a general Li ghtweight S emi-supervised T ext classification framework (LiST), which significantly enhances the performance of lightweight models in semi-supervised settings, thus improving inference efficiency. LiST employs a teacher intervention strategy, where pseudo-labels initially rely on the teacher model and gradually shift to the lightweight model’s own predictions, progressively correcting the teacher’s incorrect predictions over time. Experimental results on multiple benchmark datasets demonstrate the generality and effectiveness of LiST. It not only approaches or exceeds the performance of the teacher model (UDA) at a 20 × faster inference speed but also outperforms existing lightweight methods in both performance and efficiency. Notably, LiST achieves significant performance improvements even with extremely few labeled samples, such as a 36.4% increase in accuracy on AG News with only 2 labeled samples per class.}, archive = {J_ASOC}, author = {Shaohuan Cheng and Wenyu Chen and Wanlong Liu and Hong Qu}, doi = {10.1016/j.asoc.2025.113844}, journal = {Applied Soft Computing}, month = {12}, pages = {113844}, shortjournal = {Appl. Soft. Comput.}, title = {Improving lightweight semi-supervised text classification via teacher intervention}, volume = {184}, year = {2025}, }
• (2025). Chaotic property based multi-interval informer modeling method for long-term photovoltaic power generation prediction. ASOC, 184, 113843. (www)

  @article{ , abstract = {The long-term prediction of photovoltaic (PV) power generation capacity can significantly enhance the maintenance planning of photovoltaic power stations and support the long-term development strategies of power supply and distribution networks. Currently, two major challenges hinder the realization of effective long-term prediction for PV power generation: First, PV power generation data is heavily influenced by environmental factors, leading to high randomness and significant volatility in long-term data sequences; Second, the temporal continuity of long-term data must be considered, which complicates the construction of an accurate predictive model. To address these issues, this paper proposes a long-term PV power generation prediction method based on interval division, modeled using an improved Informer architecture with a modified activation function. Initially, by analyzing the fluctuation characteristics of PV power generation data, the K-nearest neighbors algorithm is applied for data interpolation, while the moving average (MA) method is employed for data smoothing, effectively reducing data randomness while preserving the overall trend. Subsequently, the Kmeans＋＋ clustering algorithm is utilized to group the generation data into multiple intervals, thereby enhancing feature similarity within each cluster. Finally, the activation function of the Informer model is replaced with ReLU to improve its adaptability to photovoltaic data characteristics. Additionally, this study introduces evaluation metrics such as the Hurst exponent and the Maximum Lyapunov Exponent (MLE) to assess the predictability of partitioned data and identify the interval to which predicted values belong. Simulation experiments demonstrate that, compared to several commonly used prediction models, the proposed interval-based approach achieves superior performance in long-term PV power generation prediction.}, archive = {J_ASOC}, author = {Ying Han and Xinggang Hu and Kun Li}, doi = {10.1016/j.asoc.2025.113843}, journal = {Applied Soft Computing}, month = {12}, pages = {113843}, shortjournal = {Appl. Soft. Comput.}, title = {Chaotic property based multi-interval informer modeling method for long-term photovoltaic power generation prediction}, volume = {184}, year = {2025}, }
• (2025). A noisy multi-objective evolutionary optimization algorithm based on elman neural network. ASOC, 184, 113842. (www)

  @article{ , abstract = {Neural networks have emerged as potent tools for tackling multi-objective optimization problems with remarkable success. This research delves into the performance of dynamic neural networks in resolving noisy multi-objective optimization problems (NMOPs). The local regression structure of the Elman neural network enables it to deal with the dynamic noise problem well. This paper integrates the Elman neural network into the framework of the non-dominated sorting genetic algorithm II (NSGA-II) to solve NMOP, called E-NSGA-II. In this method, the Elman neural network is employed for modeling, estimating the fitness value of each individual. Simultaneously, a hybrid selection solution strategy is implemented to select individuals for modeling, providing more diverse solutions in the modeling process to improve the convergence. Furthermore, E-NSGA-II incorporates a noise-driven sampling mechanism that intelligently adapts the sampling frequency based on the noise intensity. This feature could not only enhance the accuracy of neural network modeling, but also minimize the amount of calculation. Notably, the embedding of neural network does not impose much additional evaluation overheads, thereby bolstering the overall efficiency of E-NSGA-II. Experimental results prove the superiority of E-NSGA-II over four state-of-the-art noisy multi-objective optimization algorithms on dealing with NMOPs.}, archive = {J_ASOC}, author = {Jianxia Li and Ruochen Liu and Wanfeng Chen and Weibin Li}, doi = {10.1016/j.asoc.2025.113842}, journal = {Applied Soft Computing}, month = {12}, pages = {113842}, shortjournal = {Appl. Soft. Comput.}, title = {A noisy multi-objective evolutionary optimization algorithm based on elman neural network}, volume = {184}, year = {2025}, }
• (2025). An effective detection algorithm for small UAV based on lightweight you-only-look-once (YOLOv4-l) approach. ASOC, 184, 113841. (www)

  @article{ , abstract = {The control and monitoring of small Unmanned Aerial Vehicles (UAV) plays a crucial role in national defense and security. However, due to their compact size and high mobility, the detection of small UAV across diverse scenarios remains a significant challenge. To address this issue, this study proposes an improved detection algorithm tailored for small UAV. The model is initially trained on a virtual dataset, and the learned parameters are transferred to real-world data through a transfer learning framework. To optimize anchor box generation, clustering analysis is performed on bounding box dimensions, resulting in anchor boxes with appropriate scales and aspect ratios. Furthermore, the Ghost module is introduced to replace conventional convolutions in CSPDarknet53, enhancing feature extraction efficiency. An Efficient Channel Attention (ECA) mechanism is also incorporated to strengthen output feature representations and improve the capture of texture details critical for small target detection. Through experiments, the proposed algorithm can achieve the mAP0.5 of 82.2 %. Experimental results demonstrate the effectiveness of the proposed small UAV detection method.}, archive = {J_ASOC}, author = {Guoning Li and Jianghao Cheng and Yanyan Liu and Jin Li and Zengming Lv and Qiang Li}, doi = {10.1016/j.asoc.2025.113841}, journal = {Applied Soft Computing}, month = {12}, pages = {113841}, shortjournal = {Appl. Soft. Comput.}, title = {An effective detection algorithm for small UAV based on lightweight you-only-look-once (YOLOv4-l) approach}, volume = {184}, year = {2025}, }
• (2025). TPLLM: A traffic prediction framework based on pretrained large language models. ASOC, 184, 113840. (www)

  @article{ , abstract = {Traffic prediction constitutes a critical component in sustainable urban data analysis, playing a pivotal role in optimizing transportation systems for reduced carbon emissions and improved energy efficiency. The precision of prevailing deep learning-driven traffic prediction models typically improves as the volume of training data increases. However, the procurement of comprehensive spatiotemporal datasets for traffic is often fraught with challenges, primarily stemming from the substantial costs associated with data collection and retention. This limitation severely hinders the deployment of models in regions with insufficient historical data. Consequently, developing a model that can achieve accurate predictions and good generalization ability in areas with limited historical traffic data is a challenging problem. It is noteworthy that the rapidly advancing pretrained Large Language Models (LLMs) of recent years demonstrate exceptional proficiency in cross-modality knowledge transfer and few-shot learning. Recognizing the sequential nature of traffic data, similar to language, we introduce TPLLM, a novel traffic prediction framework leveraging LLMs. In this framework, we construct a sequence embedding layer based on Convolutional Neural Networks (CNNs) and a graph embedding layer based on Graph Convolutional Networks (GCNs) to extract sequence features and spatial features, respectively. These are subsequently integrated to form inputs that are suitable for LLMs. A Low-Rank Adaptation (LoRA) fine-tuning approach is applied to TPLLM, thereby facilitating efficient learning and minimizing computational demands. Experiments on two real-world datasets demonstrate that TPLLM exhibits commendable performance in both full-sample and few-shot prediction scenarios.}, archive = {J_ASOC}, author = {Tian Ma and Yixuan Zhao and Minda Li and Yue Chen and Fangshu Lei and Yanan Zhao and Maazen Alsabaan}, doi = {10.1016/j.asoc.2025.113840}, journal = {Applied Soft Computing}, month = {12}, pages = {113840}, shortjournal = {Appl. Soft. Comput.}, title = {TPLLM: A traffic prediction framework based on pretrained large language models}, volume = {184}, year = {2025}, }
• (2025). Dynamic-linguistic fusion for brain tumor classification: A cross-modal attention framework with clinical interpretability. ASOC, 184, 113839. (www)

  @article{ , abstract = {Accurate classification of brain tumors is a key challenge in medical image analysis, and existing methods mainly rely on static MRI images, which are difficult to capture the dynamic evolutionary features of tumors. In addition, the lack of descriptive clinical text and efficient multimodal feature fusion limits the classification accuracy and generalization. To overcome these limitations, this paper proposes a new dynamic language fusion (DLF) framework. The framework (1) utilizes ResNet18 in conjunction with LSTM for time series modeling to capture the temporal evolution of tumor morphology, (2) uses BioGPT and BERT for clinical text processing for semantic understanding, and (3) applies an interpretable cross-modal attentional mechanism for feature fusion to optimize dynamic perception and semantic alignment. Experiments on 10,287 images (from four publicly available datasets) show that the proposed framework achieves an overall accuracy of 98.96 %, a precision of 99.58 %, and an AUC higher than 0.998 for all categories on the test set, which is significantly better than the existing SOTA models, and especially exhibits stronger robustness and discriminative ability in boundary ambiguity and feature overlap samples. This study validates the synergistic effect of temporal modeling and semantic understanding in brain tumor diagnosis, providing clinicians with interpretable classification outputs to assist in decision-making for complex cases, while establishing a scalable framework for medical AI systems based on large language models.}, archive = {J_ASOC}, author = {Jiancong Fan and Fangyuan Chen and Yang Li and Jiehan Zhou}, doi = {10.1016/j.asoc.2025.113839}, journal = {Applied Soft Computing}, month = {12}, pages = {113839}, shortjournal = {Appl. Soft. Comput.}, title = {Dynamic-linguistic fusion for brain tumor classification: A cross-modal attention framework with clinical interpretability}, volume = {184}, year = {2025}, }
• (2025). Path planning for flexible needle puncture based on multi-objective particle swarm optimization. ASOC, 184, 113838. (www)

  @article{ , abstract = {This paper presents a mixed-parameter MOPSO algorithm designed to address the puncture problem of flexible needles in obstacle environments. The algorithm incorporates the damage caused by the pivotal angle to soft tissues as an objective function, marking the first time this has been applied to the MOPSO algorithm. In comparison with four classical algorithms through simulation experiments, the path deviation is reduced to just 0.1 mm, significantly lower than the CPSO algorithm. The final path score achieves 35 points, surpassing the performance of other algorithms. A self-built FPAA hybrid control platform was employed for puncture experiments using a gelatin prosthesis. The experimental results confirm that the flexible needle successfully avoids obstacles and reaches the target, demonstrating the feasibility of the proposed puncture algorithm.}, archive = {J_ASOC}, author = {Ting Yang and Beibei Liu and Ru Sun and Bi Chen and Guijuan Ji}, doi = {10.1016/j.asoc.2025.113838}, journal = {Applied Soft Computing}, month = {12}, pages = {113838}, shortjournal = {Appl. Soft. Comput.}, title = {Path planning for flexible needle puncture based on multi-objective particle swarm optimization}, volume = {184}, year = {2025}, }
• (2025). A multi-goal reinforcement learning framework for motion planning of a quadrotor UAV in 3D cluttered environment with unseen random goals. ASOC, 184, 113836. (www)

  @article{ , abstract = {The standard reinforcement learning (RL) approaches render an optimal solution for a specific task by learning an optimal policy. Therefore, these RL approaches are suitable for single goal problems only. However, there exist numerous practical challenges that involve multiple goals to be achieved in a specific environment. For example, navigation of an unmanned aerial vehicle (UAV) to achieve multiple random targets in a specific three-dimensional (3D) cluttered space is a multi-goal problem. Using the standard RL approaches, the UAV agent needs to learn a separate optimal policy for each target. Thus, a multi-goal problem with random unseen goal allocations, especially in the 3D space of UAVs, increases computational effort substantially. To solve this issue, this paper presents a relatively generalized approach that learns the optimal state values and related optimal policies considering various regions of the robot environment. The proposed approach transforms the 3D robot environment into a Markov decision process (MDP) and further divides it into various virtual sub-spaces. For every initial goal position in a sub-space, value iteration algorithm enables the aerial agent to learn the optimal state values and the relevant policy. These optimal state values and policies are then stored in a replay buffer for later use. Once the initial state values for every sub-space are learned, the proposed approach allows the agent to use them from replay buffer and run a local search algorithm to connect every new goal position to any feasible state in the existing state values. Our proposed framework significantly reduces the computational effort for multiple unseen goal targets by restricting the re-computation of state values for each new goal. To validate the proposed method, a simulator is designed and an RL-based motion planning approach for a quadrotor UAV is presented. The results under various scenarios confer the superior performance of our anticipated multi-goal RL framework.}, archive = {J_ASOC}, author = {Ghulam Farid and Lanyong Zhang and Talha Younas and Muhammad Ilyas and Asma Iqbal}, doi = {10.1016/j.asoc.2025.113836}, journal = {Applied Soft Computing}, month = {12}, pages = {113836}, shortjournal = {Appl. Soft. Comput.}, title = {A multi-goal reinforcement learning framework for motion planning of a quadrotor UAV in 3D cluttered environment with unseen random goals}, volume = {184}, year = {2025}, }
• (2025). Wind speed prediction based on U-shaped 2D multi-scale model. ASOC, 184, 113835. (www)

  @article{ , abstract = {Accurate wind speed forecasting is essential for enhancing energy efficiency and reducing maintenance costs in wind power systems. However, existing CNN-LSTM and Transformer-based models face limitations in capturing periodic features, modeling long-term dependencies, and preserving information during multi-layer encoding. To overcome these challenges, this paper proposes a novel U-shaped 2D multi-scale model that integrates Bi-RLSTM encoding, wavelet transform, and multi-scale 2D convolution. The proposed model first encodes input sequences into a high-dimensional space via feature embedding and Bi-RLSTM, effectively capturing long-term dependencies. Subsequently, a wavelet transform extracts primary fluctuation patterns and their periods, converting 1D sequences into 2D feature maps to enhance periodic feature representation while preserving temporal information. Multi-scale convolutional layers are then employed to extract fine-grained spatial–temporal features from these maps. Finally, a Bi-RLSTM decoding layer, augmented with skip connections, mitigates information loss during deep encoding and reinforces long-range correlation modeling. Extensive experiments on two real-world wind speed datasets demonstrate that the proposed model significantly outperforms CNN-LSTM and attention-based approaches achieving superior performance in terms of MAE, RMSE, and R 2 , confirming its effectiveness and application potential.}, archive = {J_ASOC}, author = {Yue Gao and Zhongda Tian}, doi = {10.1016/j.asoc.2025.113835}, journal = {Applied Soft Computing}, month = {12}, pages = {113835}, shortjournal = {Appl. Soft. Comput.}, title = {Wind speed prediction based on U-shaped 2D multi-scale model}, volume = {184}, year = {2025}, }
• (2025). Enhancing multi-task learning-based tornado identification using spatial and temporal information from weather radar images. ASOC, 184, 113834. (www)

  @article{ , abstract = {Tornadoes, as dynamic weather phenomena, exhibit unique spatial and temporal evolution characteristics that reflect their formation and development. Existing tornado detection algorithms often struggle with high false alarm rates, primarily due to insufficient capture of temporal correlations in tornado development. As an improvement, we propose a multi-task tornado identification network with three-dimensional temporal and spatial information (TS-MTINet). Taking continuous three-frame radar data as input, the Multi-frame Temporal Interaction Block (MTIB) utilizes multi-head attention to model the dynamic interaction information between the radar data, thus exploring in-depth the temporal features during tornado development. Further, we design a Spatial-Temporal Enhancement Module (STEM), which analyzes the difference information between continuous data to extract local and global spatial and temporal feature variations about tornadoes. Based on this architecture, TS-MTINet incorporates a multi-task learning framework to perform tornado detection and number estimation tasks simultaneously, thus extracting comprehensive information related to tornadoes. To validate the performance of the proposed model, we construct the first Chinese tornado identification dataset with fine radar features. The experimental results show that the proposed method shows significant advantages in several evaluation metrics, especially in reducing false alarms. In practical case studies, compared to the traditional TVS method, TS-MTINet achieves an increase in POD of approximately 30% and a decrease in FAR of about 20% in several typical tornado events. Particularly in environments with strong interference, TS-MTINet demonstrates higher detection accuracy, reflecting greater robustness and practical value.}, archive = {J_ASOC}, author = {Jinyang Xie and Kanghui Zhou and Lei Han and Liang Guan and Maoyu Wang and Yongguang Zheng and Hongjin Chen and Jiaqi Mao}, doi = {10.1016/j.asoc.2025.113834}, journal = {Applied Soft Computing}, month = {12}, pages = {113834}, shortjournal = {Appl. Soft. Comput.}, title = {Enhancing multi-task learning-based tornado identification using spatial and temporal information from weather radar images}, volume = {184}, year = {2025}, }
• (2025). Evaluating regional green industry competitiveness in china: A CoSOGR-MABAC-sort framework. ASOC, 184, 113831. (www)

  @article{ , abstract = {With the rise of concepts such as the green economy and sustainable development, the development of green industry has become increasingly urgent. Regional green industry competitiveness, as a key indicator reflecting the development status and competitive advantages of regional green industry, has not yet received sufficient attention and research. To fill this research gap, this paper proposes an evaluation framework, CoSOGR-MABAC-Sort (Combined Subjective-Objective, Grey Relational analysis-Multi-Attributive Border Approximation area Comparison-Sort), to assess the green industry competitiveness in 31 regions of China. In the proposed framework, the MABAC-Sort method is a novel multi-criteria decision sorting method with six classification rules. Compared to other MCDS methods, this method does not require predefined profile boundaries and can provide more detailed classification results. Furthermore, we propose an optimal model , called CoSOGR, to determine the weights of each indicator. Finally, by collecting subjective and objective data, we use the proposed framework to assess the green industry competitiveness in the 31 provinces of China. The main findings are as follows: 1) The CoSOGR demonstrates the highest consistency (99.33 %) in regions ranking compared to existing weight methods (CRITIC, EWM, ROCOSD, and MEREC). 2) The CoSOGR addresses the weight determination issues that ROCOSD (a mixed-integer linear programming model) cannot resolve. 3) The CoSOGR-MABAC-Sort method exhibit strong robustness and stability. 4) The green industry competitiveness in each region shows a positive correlation with the sustainable development of that region.}, archive = {J_ASOC}, author = {Jiafu Su and Baojian Xu and Lvcheng Li and Yijun Chen and Hongyu Liu and Na Zhang}, doi = {10.1016/j.asoc.2025.113831}, journal = {Applied Soft Computing}, month = {12}, pages = {113831}, shortjournal = {Appl. Soft. Comput.}, title = {Evaluating regional green industry competitiveness in china: A CoSOGR-MABAC-sort framework}, volume = {184}, year = {2025}, }
• (2025). AutoTandemML: Active learning enhanced tandem neural networks for inverse design problems. ASOC, 184, 113828. (www)

  @article{ , abstract = {Inverse design in science and engineering involves determining optimal design parameters that achieve desired performance outcomes, a process often hindered by the complexity and high dimensionality of design spaces, leading to significant computational costs. To tackle this challenge, we propose a novel hybrid approach that combines active learning with Tandem Neural Networks to enhance the efficiency and effectiveness of solving inverse design problems. Active learning allows to selectively sample the most informative data points, reducing the required dataset size without compromising accuracy. We investigate this approach using three benchmark problems: airfoil inverse design, photonic surface inverse design, and scalar boundary condition reconstruction in diffusion partial differential equations. We demonstrate that integrating active learning with Tandem Neural Networks outperforms standard approaches across the benchmark suite, achieving better accuracy with fewer training samples.}, archive = {J_ASOC}, author = {Luka Grbcic and Juliane Müller and Wibe Albert de Jong}, doi = {10.1016/j.asoc.2025.113828}, journal = {Applied Soft Computing}, month = {12}, pages = {113828}, shortjournal = {Appl. Soft. Comput.}, title = {AutoTandemML: Active learning enhanced tandem neural networks for inverse design problems}, volume = {184}, year = {2025}, }
• (2025). Personalized cross-domain recommendation with meta networks and contrastive learning. ASOC, 184, 113827. (www)

  @article{ , abstract = {In this paper, we propose a novel cross-domain recommendation (CDR) model that integrates matrix factorization (MF), attention networks, meta-networks, and contrastive learning (CL) to address the challenges of data sparsity and cold-start problems in recommender systems. Our model consists of five modules: the MF module extracts user and item embeddings from rating matrices in both information-rich and information-scarce domains; the transferable feature extraction module uses an attention network to identify and extract transferable features from users’ interactions in the information-rich domain; the cross-domain meta-knowledge transfer module employs a meta-network to transfer these features across domains while preserving users’ personalized preferences; the intra-domain CL module ensures temporal consistency of users’ preferences by learning from their interaction sequences in the information-rich domain; and the inter-domain CL module leverages feedback from users’ interactions in the information-scarce domain to refine the transferable features. We conduct extensive experiments on the Amazon and Douban datasets, evaluating our model across three different CDR tasks. The results demonstrate that our proposed model outperforms other prevalent CDR models in terms of MAE and RMSE. Additionally, our model shows the robust performance in cold-start scenarios, effectively utilizing both intradomain and interdomain knowledge to enhance the recommendation accuracy.}, archive = {J_ASOC}, author = {Shudong Liu and Xiping Hao and Xu Chen and Wenming Ma and Feng Gu}, doi = {10.1016/j.asoc.2025.113827}, journal = {Applied Soft Computing}, month = {12}, pages = {113827}, shortjournal = {Appl. Soft. Comput.}, title = {Personalized cross-domain recommendation with meta networks and contrastive learning}, volume = {184}, year = {2025}, }
• (2025). Efficient real-time visual anomaly detection via frequency-aware diffusion model and information fusion. ASOC, 184, 113826. (www)

  @article{ , abstract = {Real-time visual anomaly detection is always a big challenge since advanced techniques usually suffer from intensive computation, while simple methods cannot obtain desirable performance. To address this issue, an Efficient Real-time Anomaly Detection method via frequency-aware diffusion model and information fusion (ERAD) is proposed, which uses the diffusion model to achieve a nice reconstruction for the frequency information obtained by Discrete Wavelet Transform (DWT) and utilizes the information fusion technique to take full consideration of local and global various features. First, the input image is fed into the DWT module to produce one low-frequency and three high-frequency coefficient images, in this way, data dimensions can be reduced by 75% but the important information can be greatly retained. Then, a diffusion model with one-step denoising is developed rather than the traditional iterative denoising to accelerate the reconstruction speed. As well, information fusion in the framework of selective fusion is embedded into the reconstruction process to improve the model performance. Furthermore, during the segmentation, a wavelet-based upsampling module is put forward to seamlessly combine global semantic context with detailed edge information, achieving both consistent semantics and high-resolution feature reconstruction. Finally, extensive experiments conducted on a variety of benchmark datasets demonstrate the remarkable superiority of our method in both model performance and time efficiency. Specifically, the proposed method achieves an Image AUROC of 99.7 and a Pixel AUROC of 99.5, with an inference time of only 0.04 s.}, archive = {J_ASOC}, author = {Xianzhe Yao and Ping Kong and Quanquan Li and Yan Song}, doi = {10.1016/j.asoc.2025.113826}, journal = {Applied Soft Computing}, month = {12}, pages = {113826}, shortjournal = {Appl. Soft. Comput.}, title = {Efficient real-time visual anomaly detection via frequency-aware diffusion model and information fusion}, volume = {184}, year = {2025}, }
• (2025). Dual-splitting conformal prediction for multi-step time series forecasting. ASOC, 184, 113825. (www)

  @article{ , abstract = {Time series forecasting is crucial for applications like resource scheduling and risk management, where multi-step predictions provide a comprehensive view of future trends. Uncertainty Quantification (UQ) is a mainstream approach for addressing forecasting uncertainties, with Conformal Prediction (CP) gaining attention due to its model-agnostic nature and statistical guarantees. However, most variants of CP are designed for single-step predictions and face challenges in multi-step scenarios, such as reliance on real-time data and limited scalability. This highlights the need for CP methods specifically tailored to multi-step forecasting. We propose the Dual-Splitting Conformal Prediction (DSCP) method, a novel CP approach designed to capture inherent dependencies within time-series data for multi-step forecasting. Experimental results on real-world datasets from four different domains demonstrate that DSCP significantly outperforms existing CP variants in terms of the Winkler Score, improving performance by up to 23.59% compared to state-of-the-art methods. Furthermore, the deployment of DSCP for renewable energy generation and IT load forecasting in the power management of a real-world trajectory-based application achieves an 11.25% reduction in carbon emissions through predictive optimization of data center operations and control strategies.}, archive = {J_ASOC}, author = {Qingdi Yu and Zhiwei Cao and Ruihang Wang and Zhen Yang and Lijun Deng and Min Hu and Yong Luo and Xin Zhou}, doi = {10.1016/j.asoc.2025.113825}, journal = {Applied Soft Computing}, month = {12}, pages = {113825}, shortjournal = {Appl. Soft. Comput.}, title = {Dual-splitting conformal prediction for multi-step time series forecasting}, volume = {184}, year = {2025}, }
• (2025). FDP-net: Fourier transform guided lightweight depthwise and pointwise dynamic pooling based neural network for medical image classification. ASOC, 184, 113824. (www)

  @article{ , abstract = {Nowadays, deep learning-based medical image classification has become essential, especially in developing countries because of the high volume of patients with less medical professionals as well as required infrastructures. Deep learning models often help in the early detection of diseases; however, it require a high amount of processing power, and sometimes it becomes less scalable for various computer-aided diagnosis. To this end, in this paper, a lightweight Fourier Transform guided Depth and Pointwise Dynamic Pooling based Neural Network (FDP-Net), has been proposed for medical image classification. This paper introduces a Depth and Pointwise Feature Fusion (DPFF) block for learning the important features with less computation and without increasing the model parameters. It also proposes a dynamic pooling technique, an alternative to traditional max-pooling, which dynamically selects the important features. The proposed FDP-Net model is trained to classify medical images with the guidance of Fourier Transformation and multitask loss function, which makes the model converge faster and reduces overfitting. The proposed model has been tested on Acute Lymphoblastic Leukemia (ALL) dataset, Peripheral Blood Cell (PBC) dataset, and Raabin White blood Cell (Raabin-WBC) dataset, and it outperforms the state-of-the-art models with 100%, 98.13% and 96.79% classification accuracies, respectively. Additionally, the proposed model is made with only 0.349 million parameters, thereby enabling faster processing. Code will be avilabe at https://github.com/asfakali/FDP-Net .}, archive = {J_ASOC}, author = {Asfak Ali and Rajdeep Pal and Aishik Paul and Ram Sarkar}, doi = {10.1016/j.asoc.2025.113824}, journal = {Applied Soft Computing}, month = {12}, pages = {113824}, shortjournal = {Appl. Soft. Comput.}, title = {FDP-net: Fourier transform guided lightweight depthwise and pointwise dynamic pooling based neural network for medical image classification}, volume = {184}, year = {2025}, }
• (2025). Explicit cooperation mechanism and multimodal fusion prediction model empower DouDizhu agents. ASOC, 184, 113823. (www)

  @article{ , abstract = {Imperfect information games require agents to make decisions under uncertainty, presenting significant challenges yet offering broad applicability in real-world scenarios. DouDizhu, a representative example, features complex cooperative-competitive dynamics, imperfect information, and a large strategy space. To address these challenges, a strategic decision-making algorithm is developed by integrating multi-agent credit allocation with information prediction. Specifically, an explicit credit allocation mechanism based on team reward decomposition improves cooperative behavior among peasant agents by enabling more stable and targeted policy updates. Meanwhile, combining long short-term memory networks and multi-head attention enhances the prediction of hidden opponent information through multimodal feature fusion. Moreover, convolutional neural networks are incorporated into the DouZero framework to extract high-level features and reduce the policy solution space. The resulting CAPRE_DMC significantly improves agent performance in adversarial settings. Extensive evaluations demonstrate that CAPRE_DMC outperforms baseline DouDizhu agents, achieving substantial gains in point margin and win percentage. Additionally, evaluations on benchmark multi-agent cooperative tasks demonstrate the framework’s scalability and generality in large-scale imperfect-information environments.}, archive = {J_ASOC}, author = {Jiao Wang and Longyue Fu and Xiang Li and Hongchen Luo and Zhifen Guo}, doi = {10.1016/j.asoc.2025.113823}, journal = {Applied Soft Computing}, month = {12}, pages = {113823}, shortjournal = {Appl. Soft. Comput.}, title = {Explicit cooperation mechanism and multimodal fusion prediction model empower DouDizhu agents}, volume = {184}, year = {2025}, }
• (2025). A lightweight detection method for X-ray security inspection based on YOLOV8. ASOC, 184, 113822. (www)

  @article{ , abstract = {In the field of public safety, the prohibited items detection during X-ray security inspection is crucial for preventing potential threats. The recent advances in artificial intelligence, particularly deep learning, have achieved success in this area. However, the high number of parameters and computational load of these deep learning-based object detection methods result in significant hardware requirements limiting their practical application. To address these issues, a lightweight method for detecting prohibited items in X-ray security inspections, utilizing the YOLOV8 framework, is proposed in this paper. The key innovations of this method are threefold. First, a lightweight convolution module is designed to optimize the YOLOV8 model. This optimization significantly reduces both the number of parameters and the computational load, making the model more efficient. Second, an adaptive spatial-and-channel attention module is designed to enhance feature extraction capabilities. This module enhances feature extraction capabilities without compromising detection accuracy. Third, the Wise-IOU loss function is incorporated to enhance the overall performance of the detection method during training. Finally, the method is evaluated on our real X-ray pseudo-color image dataset, PIDRAY and CLCXRAY with the existing methods. Our proposed method attains a detection accuracy of 98.83 % on the self-constructed dataset, demonstrating a 0.6 % improvement, while concomitantly reducing the parameters by 64.2 % and computational load by 66.3 %. Furthermore, it achieves accuracy enhancements of 0.7 % and 0.82 % on the two public datasets, respectively. The experimental results indicate that this method not only reduces hardware requirements due to a lower number of parameters and computational load but also broadens the model's applicability. This work underscores the importance of balancing model complexity with performance and sets the stage for future research in AI-driven security inspection technologies.}, archive = {J_ASOC}, author = {Xizhuo Yu and Chunyang Chen and Shu Cheng and Jingming Li}, doi = {10.1016/j.asoc.2025.113822}, journal = {Applied Soft Computing}, month = {12}, pages = {113822}, shortjournal = {Appl. Soft. Comput.}, title = {A lightweight detection method for X-ray security inspection based on YOLOV8}, volume = {184}, year = {2025}, }
• (2025). Probabilistic transformation of basic probability assignment based on weighted visibility graph networks. ASOC, 184, 113821. (www)

  @article{ , abstract = {Dempster-Shafer evidence theory (DSET) provides a powerful framework for uncertain reasoning, offering a theoretical basis for decision-making under ambiguity. However, its core representation, the basic probability assignment (BPA), cannot be directly applied to probabilistic decision-making, prompting the need for effective probability transformation methods. A key challenge lies in quantifying the uncertainty inherent in BPAs to guide this transformation process. To address this, we propose an improved probabilistic transformation method that integrates belief entropy and weighted visibility graph networks, which yields more accurate and interpretable probability distributions than existing approaches. Specifically, given a frame of discernment and mass function, we first apply two refined belief entropy measures to evaluate the informational content of each focal element. Based on these entropy-derived orderings, we construct a weighted visibility graph that captures the structural relationships among focal elements. The weights from this graph are then used to compute a proportional belief transformation. Experimental validation across benchmark cases on classical BPA scenarios demonstrates that our method outperforms traditional approaches in terms of entropy consistency and decision quality, as evidenced by lower Kullback–Leibler (KL) divergence, higher probability information capacity (PIC), and reduced Shannon entropy. These results highlight the method’s dual advantage in balancing decisiveness (via PIC maximization) and fidelity (via entropy minimization), making it a robust tool for uncertainty-aware decision support systems.}, archive = {J_ASOC}, author = {Yongchuan Tang and Kangkang Wu and Rongfei Li and He Guan and Deyun Zhou and Yubo Huang}, doi = {10.1016/j.asoc.2025.113821}, journal = {Applied Soft Computing}, month = {12}, pages = {113821}, shortjournal = {Appl. Soft. Comput.}, title = {Probabilistic transformation of basic probability assignment based on weighted visibility graph networks}, volume = {184}, year = {2025}, }
• (2025). Electric network frequency-based digital audio tampering event identification using multimodal feature interaction network. ASOC, 184, 113820. (www)

  @article{ , abstract = {In the field of digital Audio Tampering Event (ATE) detection based on Electric Network Frequency (ENF), accurately extracting ENF signals is crucial for tampering event identification. However, ENF signals are susceptible to noise interference, and it is difficult to establish an effective matching relationship with the reference frequency database, which poses a challenge to the effectiveness of existing ATE detection methods. To solve these problems, a Multimodal Feature Interactive Network (MFIN) is proposed for ATE identification under low-SNR conditions. First, a Modified Kaiser-window-based S-Transform (MKST) method is proposed to extract the estimated frequency, phase, rate of change of frequency, and rate of change of phase of the ENF component in digital audio. The frequency and time resolution are increased through the improved control function, thereby enhancing the accuracy of ENF estimation. Subsequently, a Multi-Branch Multi-Scale Attention Convolutional (MBSAC) neural network is further proposed to extract and classify ENF features. In MBSAC, the Multi-Branch Multi-Scale Temporal (MBST) attention fusion block, serving as the network’s primary structure, increases feature diversity and enhances classification performance. Finally, the ATE data acquisition hardware platform is built. Experimental results on a dataset comprising six types of ATEs demonstrate that the proposed MFIN outperforms several state-of-the-art ATE detection methods in both ENF extraction and tampering type classification.}, archive = {J_ASOC}, author = {Bing Li and Junfeng Duan and Yao Zheng and Xinxin Cai and Wei Qiu and He Yin and Wenxuan Yao}, doi = {10.1016/j.asoc.2025.113820}, journal = {Applied Soft Computing}, month = {12}, pages = {113820}, shortjournal = {Appl. Soft. Comput.}, title = {Electric network frequency-based digital audio tampering event identification using multimodal feature interaction network}, volume = {184}, year = {2025}, }
• (2025). Enhancing financial resilience in manufacturing SMEs: A q-rung picture fuzzy set-based decision framework for digital transformation adoption. ASOC, 184, 113819. (www)

  @article{ , abstract = {Digital transformation (DT) has emerged as a crucial strategy for enhancing the financial resilience of small and medium-sized enterprises (SMEs) in the manufacturing sector. Despite its importance, there is a significant gap in the quantitative evaluation of the factors influencing SMEs’ DT adoption. This study addresses this gap by developing a decision framework using q-rung picture fuzzy sets (q-RPF) to identify and analyze the drivers of DT implementation aimed at boosting financial resilience in manufacturing SMEs. The proposed framework incorporates the q-RPF-weighted Heronian mean aggregation operator to aggregate expert evaluations and capture the interrelationships among input decision data. A novel weighting approach, based on the q-RPF deviation measure, is introduced to assess the significance of various enablers. Furthermore, the q-RPF-MARCOS’H model is employed to evaluate the effectiveness of DT in enhancing financial resilience across different SMEs by integrating the methodologies mentioned above. A numerical example involving manufacturing SMEs from a specific city demonstrates the practical application of the q-RPF-MARCOS’H model-based framework. The results reveal that “concurrent operations” (0.102) is a significant enabler of DT adoption in promoting financial resilience. The framework’s validity is confirmed through both sensitivity and comparative analyses. These outcomes offer recommendations for policymakers and industry leaders to design effective incentives for DT adoption and provide practical guidance for SMEs seeking to leverage DT for improved financial resilience.}, archive = {J_ASOC}, author = {Zhengyan Yang and Wei Zhong Wang and Zelin Wang and Muhammet Deveci and Dursun Delen}, doi = {10.1016/j.asoc.2025.113819}, journal = {Applied Soft Computing}, month = {12}, pages = {113819}, shortjournal = {Appl. Soft. Comput.}, title = {Enhancing financial resilience in manufacturing SMEs: A q-rung picture fuzzy set-based decision framework for digital transformation adoption}, volume = {184}, year = {2025}, }
• (2025). A regret theory-based three-way decision model under comparative linguistic expressions. ASOC, 184, 113816. (www)

  @article{ , abstract = {In real world, experts often prefer to utilize linguistic expressions over numerical data when evaluating alternatives. However, due to the complexity of actual decision-making, single linguistic terms are insufficient for experts to express their judgments accurately. Thus, it is necessary to employ a richer form of linguistic expression, known as comparative linguistic expressions (CLEs). Furthermore, existing multi-attribute decision-making (MADM) models under CLE suffer from the following limitations: (1) they fail to provide decision-making references for alternatives; (2) they do not incorporate the psychological factors of experts. In order to address the aforementioned challenges, this paper proposes a novel regret theory (RT)-based three-way decision (TWD) model under CLE. First, the attribute weights are calculated by an improved optimization model. This model combines index variability with comprehensive entropy, enabling a more objective conclusion to be drawn regarding the relative importance of attributes. Second, an enhanced neighborhood relationship is introduced, which is shown to fulfill the properties of Symmetry, Reflexivity, and Non-transitivity. Building on this foundation, this study integrates RT with the neighborhood relationship to construct a wide TWD framework. This framework incorporates decision-making references for the alternatives and accounts for the influence of experts’ psychological factors. Subsequently, the ranking of alternatives is determined using the technique for order preference by similarity to ideal solution (TOPSIS) method. Finally, the feasibility of the proposed method is demonstrated through a real-case study. Comparative experiments and parameter sensitivity analysis are designed to demonstrate the superiority and effectiveness of the model.}, archive = {J_ASOC}, author = {Zhanhao Liu and Huangjian Yi and Yushan Yao and Jiajia Wang}, doi = {10.1016/j.asoc.2025.113816}, journal = {Applied Soft Computing}, month = {12}, pages = {113816}, shortjournal = {Appl. Soft. Comput.}, title = {A regret theory-based three-way decision model under comparative linguistic expressions}, volume = {184}, year = {2025}, }
• (2025). A Q-learning based estimation of distribution algorithm for automated guided vehicle scheduling in disassembly workshop. ASOC, 184, 113815. (www)

  @article{ , abstract = {With the rapid upgrade of electronic products, the disassembly process is becoming increasingly crucial in facilitating resources recycling. In practical disassembly workshops, the automated guided vehicle (AGV) plays an important role in improving efficiency of product transportation between different disassembly machines. In this paper, a Q-learning based estimation of distribution algorithm (QEDA) is proposed to solve the AGV scheduling problem in disassembly workshops with minimization of makespan for all transportation tasks. Firstly, a mathematical model for the problem is established, and specific encoding and decoding rules are designed for solution representation. Secondly, a novel update mechanism of EDA is designed by fully information utilization of both elite and poor solutions to accelerate convergence. Thirdly, a local search strategy based on Q-learning with three time-related states is designed for the elite solutions to enhance exploitation. Finally, comparative experiments on 120 benchmark test sets demonstrate that the QEDA outperforms the state-of-the-art algorithms in both solution quality and convergence speed, confirming its superior effectiveness for AGV scheduling in disassembly workshops.}, archive = {J_ASOC}, author = {Honggui Han and Teng Wang and Jingjing Wang}, doi = {10.1016/j.asoc.2025.113815}, journal = {Applied Soft Computing}, month = {12}, pages = {113815}, shortjournal = {Appl. Soft. Comput.}, title = {A Q-learning based estimation of distribution algorithm for automated guided vehicle scheduling in disassembly workshop}, volume = {184}, year = {2025}, }
• (2025). Criterion-guided polypharmacy side effects prediction with dual-view contrastive learning. ASOC, 184, 113814. (www)

  @article{ , abstract = {Predicting side effects from taking multiple drugs (polypharmacy) is a critical challenge in healthcare, with significant implications for patient safety. Traditional machine learning methods often overlook the multi-scale nature of drug information and the interdependencies among side effects, limiting their predictive power. To address these gaps, we propose CASE (Criterion-guided polyphArmacy Side Effects prediction with dual-view contrastive learning). CASE captures both microscopic (molecular graph) and macroscopic (biochemical knowledge) drug features via an adaptive substructure encoder and a biochemical feature aggregator, respectively, under a dual-view framework. Then, contrastive learning is employed to jointly balance and integrate structural and biochemical representations, enabling a more comprehensive and synergistic understanding of drug interactions. Moreover, a criterion-guided decoding strategy is designed to model complex side effect relationships. This effective design enables the proposed CASE model to achieve superior performance, with an accuracy of 88.70 %, precision of 83.62 %, recall of 96.45 %, F1-score of 89.56 %, AUROC of 94.75 %, and AUPRC of 91.16 %, respectively, on the benchmark dataset. Ablation experiments and case studies further confirm the robustness and practical utility of the model. These results demonstrate that CASE can effectively assist clinical decision-making by identifying high-risk drug combinations and delivering reliable polypharmacy risk assessments, thereby supporting safer and more personalized treatment strategies.}, archive = {J_ASOC}, author = {Yike Wang and Huifang Ma and Zihao Gao and Zhixin Li and Liang Chang}, doi = {10.1016/j.asoc.2025.113814}, journal = {Applied Soft Computing}, month = {12}, pages = {113814}, shortjournal = {Appl. Soft. Comput.}, title = {Criterion-guided polypharmacy side effects prediction with dual-view contrastive learning}, volume = {184}, year = {2025}, }
• (2025). Identifying suicidal ideations from social media posts using deep learning and explainable AI-driven approach. ASOC, 184, 113813. (www)

  @article{ , abstract = {Background At present, suicide has become one of the leading causes of unnatural deaths worldwide. Individuals having suicidal urges often express their self-harming ideas through social media posts. Early identification of such ideations is critical for timely intervention and prevention. Besides, continuous assessment of the texts containing suicidal thoughts can uncover the hidden triggers of suicidal urges. This study presents a comprehensive approach to analyze the user-generated textual contents on social media that reflect suicidal ideas. Methodology For identifying the underlying topics that express suicidal ideations, this study has employed the Latent Dirichlet Allocation (LDA) model. Semantic Network Analysis (SNA) is used to gain a deeper quantitative and qualitative insight into these texts. Besides, an exploratory investigation of different deep learning (DL) models has been performed to identify the posts speculating suicidal ideations. Furthermore, this study has integrated Explainable AI (XAI) techniques like Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP) to enhance interpretability of the decisions taken by the DL models. Techniques like LDA and SNA offer a better understanding of the linguistic features of the suicidal posts, while the integration of the XAI techniques with the DL models elevates the transparency of their decisions. Contributions This study has developed an end-to-end web application, that can perform real-time classification of posts for suicidal ideation. Moreover, this application can provide insights into the rationale behind the taken decisions. This study aims to contribute to suicide prevention efforts through an innovative combination of computational techniques and AI-driven tools.}, archive = {J_ASOC}, author = {Md. Sabab Zulfiker and Nasrin Kabir and Al Amin Biswas and Md. Mashih Ibn Yasin Adan and Mohammad Shorif Uddin}, doi = {10.1016/j.asoc.2025.113813}, journal = {Applied Soft Computing}, month = {12}, pages = {113813}, shortjournal = {Appl. Soft. Comput.}, title = {Identifying suicidal ideations from social media posts using deep learning and explainable AI-driven approach}, volume = {184}, year = {2025}, }
• (2025). A lightweight network enhanced by attention-guided cross-scale interaction for underwater object detection. ASOC, 184, 113811. (www)

  @article{ , abstract = {To address the image quality degradation caused by multipath effects and scattering in underwater environments, we propose a lightweight neural network architecture, PRCII-Net, optimized for small target detection under complex underwater conditions. First, a Progressive Re-parameterized Attention-based Intra-scale Feature Interaction module (PR-AIFI) is proposed, which improves the network performance while reducing the difficulty of training and maintaining training stability. Second, a feature pyramid network named as the Cross-scale Information Interaction Feature Pyramid Network (CII-FPN) is proposed, including the two main fusion structures. The CII-FPN not only fully utilizes shallow and deep information, but also enhances the network’s spatial representation and the interaction between deep feature layers, thereby boosting the detection capability for small targets. Meanwhile, to reduce the model’s size and resource consumption, 1x1 convolutions are introduced into the backbone network for efficient channel compression and cross-channel feature fusion, significantly lowering computational complexity. Experiments on the Detecting Underwater Objects (DUO) and Real-time Underwater Object Detection (RUOD) datasets demonstrate that PRCII-Net outperforms existing real-time neural network models in mAP and F1 scores while maintaining efficient inference on resource-constrained devices.}, archive = {J_ASOC}, author = {Dehua Zhang and Changcheng Yu and Zhen Li and Chunbin Qin and Ruixue Xia}, doi = {10.1016/j.asoc.2025.113811}, journal = {Applied Soft Computing}, month = {12}, pages = {113811}, shortjournal = {Appl. Soft. Comput.}, title = {A lightweight network enhanced by attention-guided cross-scale interaction for underwater object detection}, volume = {184}, year = {2025}, }
• (2025). Preferences relative ordering by frequency inclusion technique (PROFIT) for optimizing performance evaluation. ASOC, 184, 113807. (www)

  @article{ , abstract = {The aim of this investigation is to introduce the novel concept of the neutro-cardinal family of inclusion measures, a four-parameter family of fuzzy inclusion measures tailored for single-valued neutrosophic sets (SVNSs). These measures are built using a new scalar cardinality measure for SVNSs, termed weighted average cardinality, and incorporate logical operators such as single-valued neutrosophic Frank t-norms and their corresponding dual t-conorms. The versatility offered by this four-parameter family allows the variation of t-norms and t-conorms within each measure based on a fixed combination of parameters. This flexibly supports modeling various layers of human cognitive behavior, particularly the outermost states: optimism (via fuzzy Lukasiewicz t-norms and t-conorms), neutrality (via product and probabilistic sum operators), and pessimism (via min and max operators). Moreover, capturing these distinct mental states, the family allows combinations of them, enabling the representation of more complex and deeper cognitive processes. Consequently, the proposed inclusion measures serve as robust and versatile mathematical tools for designing artificial intelligence (AI) systems in multi-criteria decision making (MCDM) environments. Building upon this family, we present a new MCDM method named PROFIT (Preferences Relative Ordering by Frequency Inclusion Technique). PROFIT is designed as a simple yet effective decision-making approach that leverages the neutro-cardinal measures to support AI-based evaluation processes. It is investigated in the context of organizational management, in which it enables diligent performance evaluation. The PROFIT model’s data-driven and adaptive structure is consistent with resilience-focused initiatives like Resilient Manufacturing, ManuChain II, and Towards Resilience in Industry 5.0. As such, it serves as a strategic tool for optimizing decision-making, enhancing performance, and strengthening organizational resilience in dynamic and complex environments.}, archive = {J_ASOC}, author = {Madiha Qayyum and Dania Farooq and Muhammad Riaz and Muhammad Aslam and Dragan Pamucar}, doi = {10.1016/j.asoc.2025.113807}, journal = {Applied Soft Computing}, month = {12}, pages = {113807}, shortjournal = {Appl. Soft. Comput.}, title = {Preferences relative ordering by frequency inclusion technique (PROFIT) for optimizing performance evaluation}, volume = {184}, year = {2025}, }
• (2025). A state alignment-centric approach to federated system identification: The FedAlign framework. ASOC, 184, 113800. (www)

  @article{ , abstract = {This paper presents FedAlign, a Federated Learning (FL) framework, designed for System Identification (SYSID) of linear State-Space Models (SSMs) by aligning state representations. Local workers can learn linear SSMs with equivalent representations but different parameter basins. We demonstrate that directly aggregating these local SSMs via FedAvg results in a global model with altered system dynamics. FedAlign overcomes this problem by employing similarity transformation matrices to align state representations of local SSMs, thereby establishing a common parameter basin that retains the dynamics of local SSMs. FedAlign computes similarity transformation matrices via two distinct approaches. In FedAlign-A, we represent the global SSM in controllable canonical form (CCF). We use control theory to analytically derive similarity transformation matrices that convert each local SSM into this form. Yet, establishing global SSM in CCF brings additional alignment challenges in multi-input multi-output SYSID, as CCF representation is not unique, unlike in single-input single-output SYSID. In FedAlign-O, we address the alignment challenges by reformulating the local parameter basin alignment problem as an optimization task. We set the parameter basin of a local worker as the common parameter basin and solve least square problems to obtain the transformation matrices needed to align the remaining local SSMs. The experiments conducted on synthetic and real-world datasets show that FedAlign outperforms FedAvg, converges faster, and provides improved global SSM stability thanks to local parameter basins’ alignment.}, archive = {J_ASOC}, author = {Ertuğrul Keçeci and Müjde Güzelkaya and Tufan Kumbasar}, doi = {10.1016/j.asoc.2025.113800}, journal = {Applied Soft Computing}, month = {12}, pages = {113800}, shortjournal = {Appl. Soft. Comput.}, title = {A state alignment-centric approach to federated system identification: The FedAlign framework}, volume = {184}, year = {2025}, }
• (2025). Optimal harmonics prediction for distribution systems powered by multi-energy sources using bidirectional long-short term memory combined with data sequence. ASOC, 184, 113799. (www)

  @article{ , abstract = {A multi-energy resource aims to maintain a balance between energy output and load consumption and to ensure power continuity during different operating conditions. The harmonic distortions can be estimated from the output current of a harmonic source, which may not fully reflect its true harmonic distortions due to the interactions between the state changes at the power network level and the harmonic sources. System operators monitor each system's harmonic performance under different conditions of operation to find the actual contribution of grid-connected systems to harmonic-related issues. Development of machine learning algorithms leads to effective progress in the harmonic prediction and computation. In this paper, the combined data sequencing, and Bidirectional Long-Short Term Memory (Bi-LSTM) network has been exploited for the real-time harmonic prediction of future events in multi-energy sources. The validity of the proposed Model including the applications of ANFIS, ANNs, MLRA and LSTM is conducted on the two standard systems as IEEE 9-bus and IEEE 34-bus multi energy resources system that is associated with PV systems. The simulation results, based on climate changes of solar irradiance and ambient temperature in PV systems, demonstrate that the proposed methods can accurately forecast changes in total harmonic distortion (THD) as well as the voltage profile at the point of common coupling. The performance of Bi-LSTM, original LSTM, Machine Linear Regression (MLR), and Artificial Neural Networks (ANNs) techniques were assessed. These findings provide valuable insights. Four performance validation indices, RMSE, R-squared and MSE are considered to assess the performance of the competitive learning algorithms. The results showed that in the model IEEE 9-bus, Bi-LSTM outperformed all the applied methods as its RMSE value was 0.000019 while its MSE value was 3.61e-10 and finally, the Bi-LSTM had a higher value squared error (R 2 ) was equal 1 which indicates the effectiveness of Bi-LSTM for predicting sequential total harmonic distortion. On the other hand, in case study of IEEE 34-bus, the RMSE, MSE and R 2 are 0, 3.276e-30 and 1 using Bi-LSTM which means that the Bi-LSTM leads to the best performance validation indices compared to other competitive algorithms for the tested multi-energy systems.}, archive = {J_ASOC}, author = {Hasnaa M. El-Arwash and Almoataz Y. Abdelaziz and Ragab A. El-Sehiemy}, doi = {10.1016/j.asoc.2025.113799}, journal = {Applied Soft Computing}, month = {12}, pages = {113799}, shortjournal = {Appl. Soft. Comput.}, title = {Optimal harmonics prediction for distribution systems powered by multi-energy sources using bidirectional long-short term memory combined with data sequence}, volume = {184}, year = {2025}, }
• (2025). Advanced robot path planning on rough terrain: A Q-learning-based multi-objective PSO algorithm. ASOC, 184, 113798. (www)

  @article{ , abstract = {Path planning remains one of the most extensively studied problems in mobile robotics. While particle swarm optimization (PSO) has been widely adopted for this task, conventional implementations exhibit critical limitations including being prone to local optima, lacking population diversity, and having low accuracy — all of which compromise both the quality and efficiency of path planning solutions. To address these challenges, this paper proposes a novel hybrid algorithm called MOQLCPSO that synergistically integrates Q-learning (QL) with crossover operators into multi-objective particle swarm optimization (MOPSO) for car-like mobile robots operating in known static rough terrain environments. The proposed framework aims to generate collision-free optimal paths characterized by minimal length and terrain roughness through three key innovations: Firstly, we implement QL-based dynamic parameter adaptation to autonomously adjust PSO’s inertia weight and learning factors during optimization, effectively enhancing convergence towards the Pareto front. Secondly, a strategic crossover operator is introduced to augment exploration capabilities and maintain population diversity throughout the optimization process. Finally, comprehensive comparative simulations against state-of-the-art alternatives demonstrate MOQLCPSO’s superior performance metrics. The experimental results reveal statistically significant improvements in search accuracy, population diversity, and optimization efficiency across multiple terrain complexity levels. Notably, when benchmarked against the most competitive contemporary algorithm, MOQLCPSO achieves path length reduction of 0.64%, 2.42%, and 3.87%, terrain roughness reduction of 6.68%, 5.24%, and 4.72% in simple, moderately complex, and highly complex terrains, respectively. These advancements highlight the algorithm’s strong potential for deployment in multi-objective optimization scenarios ranging from autonomous mobile robot navigation and smart manufacturing resource allocation to power grid dispatch operations.}, archive = {J_ASOC}, author = {Zhaoxia Duan and Yi Zhang and Zhen Shao and Zhen Xu and Zhengrong Xiang}, doi = {10.1016/j.asoc.2025.113798}, journal = {Applied Soft Computing}, month = {12}, pages = {113798}, shortjournal = {Appl. Soft. Comput.}, title = {Advanced robot path planning on rough terrain: A Q-learning-based multi-objective PSO algorithm}, volume = {184}, year = {2025}, }
• (2025). Combined variable precision fuzzy rough set and its application in medical diagnosis. ASOC, 184, 113797. (www)

  @article{ , abstract = {In data processing, variable precision fuzzy rough set plays a crucial role in removing noisy data. Different variable precision ideas have been used to solve various problems and achieve different goals. Among them, the variable precision ideas by Zhao and Yao are widely recognized by researchers. However, a single method is often insufficient to solve complex problems. In this study, we combined these two variable precision ideas and introduced a new variable precision fuzzy rough set model ( O - C-VPFR ), which includes the advantages and properties of the aforementioned two models. Next, using the concept of attribute importance in O - C-VPFR , we developed a simple, objective method for calculating attribute weights. Based on this method, we further developed a comprehensive decision-making method integrating Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) and Preference Ranking Organization Method for Enrichment Evaluations (PROMETHEE). The new method was applied to the problem of health risk assessment of pregnant women (HRPW, https://archive.ics.uci.edu/dataset/863/maternal+health+risk ). We conducted parameter and comparative analyses by randomly selecting 10 data points from HRPW, proving the stability and reliability of our method through the Pearson correlation coefficient. Further, we used the entire dataset and performed an ordered similarity experiment and a hypothesis testing experiment to verify the stability, effectiveness, and robustness of the proposed method. In all experiments, our method comprehensively ranked pregnant women, identified high-risk individuals, and enabled timely treatment.}, archive = {J_ASOC}, author = {Jingwen Xie and Lingqiang Li}, doi = {10.1016/j.asoc.2025.113797}, journal = {Applied Soft Computing}, month = {12}, pages = {113797}, shortjournal = {Appl. Soft. Comput.}, title = {Combined variable precision fuzzy rough set and its application in medical diagnosis}, volume = {184}, year = {2025}, }
• (2025). Optimizing evacuation path planning in high-rise building fires using an improved ant colony algorithm and dynamic window approach under smoke control scenarios. ASOC, 184, 113796. (www)

  @article{ , abstract = {Traditional emergency evacuation strategies are constrained by limited adaptability to dynamic conditions and inefficient allocation of rescue resources. To assess the impact of smoke control facilities during high-rise fire evacuations, Building EXODUS software was used to simulate occupant behavior under the influence of air curtains and mechanical smoke exhaust systems. Furthermore, to address real-time obstacle avoidance in dynamic environments and during group movement in fire emergencies, a multi-person evacuation path planning method was developed by integrating an improved ant colony optimization algorithm with the dynamic window approach. This method incorporates a heuristic function that considers fire conditions and potential fields, along with a dynamic pheromone update strategy featuring reward and punishment mechanisms. The results indicate that the effectiveness of smoke control facilities in improving evacuation outcomes ranks as follows: combined operation of mechanical smoke extraction and air curtains, mechanical smoke extraction alone, and air curtains alone. Compared to the standard ant colony algorithm, the improved algorithm reduced planned path length by 7.71 %, decreased turn times by 73.30 %, and improved computational efficiency by 6.56 %. Furthermore, under scenarios with and without operational smoke control facilities, the improved algorithm combined with the dynamic window approach reduced path overlap areas by 11.76 % and 18.19 %, respectively, and shortened path lengths by 13.49 % and 14.19 %, respectively, compared to the Building EXODUS. The proposed method effectively addresses dynamic fire environments in high-rise buildings and provides a theoretical foundation for intelligent occupant evacuation.}, archive = {J_ASOC}, author = {Yaping Yu and Qinghe Wang and Lu Jin and Ji-nan Ding and Faqi Liu}, doi = {10.1016/j.asoc.2025.113796}, journal = {Applied Soft Computing}, month = {12}, pages = {113796}, shortjournal = {Appl. Soft. Comput.}, title = {Optimizing evacuation path planning in high-rise building fires using an improved ant colony algorithm and dynamic window approach under smoke control scenarios}, volume = {184}, year = {2025}, }
• (2025). Ribonucleic-acid protein interaction prediction based on deep learning: A comprehensive survey. ASOC, 184, 113795. (www)

  @article{ , abstract = {The interaction between Ribonucleic Acids (RNAs) and proteins, also called RNA Protein Interaction (RPI), governs biological processes, including gene regulation and disease pathogenesis. This comprehensive survey examines Artificial Intelligence (AI) applications in Deep Learning-based RPI Prediction (DL-based RPIP) through eight Research Questions (RQs), analyzing 179 studies (2014–2023). The key findings include: sustained technical evolution through embryonic (2014–2017), accelerated (2018–2022), and expansion phases (2023) (RQ1); hybrid models integrating Graph Neural Networks (GNNs) (for topological interface modeling) and Transformers (for long-range dependencies) achieve state-of-the-art performance (RQ4); pretrained language models enhance small-sample learning, but the cross-species generalization declines sharply with evolutionary distance (RQ5). Critical challenges persist, including data heterogeneity across databases, the scarcity of standardized benchmarks (RQ2), and balancing the trade-off between feature encoding and information preservation (RQ3). Future advancements require biologically informed DL architectures, multi-feature fusion, and rigorous cross-validation to bridge the generalization-interpretability gap (RQ8): This would accelerate the clinical translation of predictive tools (RQ6/RQ7). As the first comprehensive analysis spanning feature encoding, modeling, evaluation, applications, and tools, this work fills a critical gap in the DL-based RPIP literature.}, archive = {J_ASOC}, author = {Danyu Li and Rubing Huang and Chenhui Cui and Dave Towey and Ling Zhou and Jinyu Tian and Bin Zou}, doi = {10.1016/j.asoc.2025.113795}, journal = {Applied Soft Computing}, month = {12}, pages = {113795}, shortjournal = {Appl. Soft. Comput.}, title = {Ribonucleic-acid protein interaction prediction based on deep learning: A comprehensive survey}, volume = {184}, year = {2025}, }
• (2025). Multi-level pooling self-adaptive evolutionary multi-objective gene selection algorithm for microarray data classification. ASOC, 184, 113794. (www)

  @article{ , abstract = {Gene selection is a vital preprocessing technique for cancer classification using microarray data, focusing on identifying a subset of genes that achieve high classification accuracy while minimizing the gene selection rate. Despite the prevalence of multi-objective optimization algorithms in microarray data classification, balancing high classification accuracy and effective guidance in microarray classification remains challenging for most existing algorithms. This study proposes a novel multi-objective gene selection method, MOGS-MLPSAE, which utilizes multi-level pooling and self-adaptive evolutionary techniques to enhance classification accuracy. MOGS-MLPSAE employs a Pareto-based ranking pool division strategy to facilitate cross-level learning among individuals and introduces a population-biased evolutionary mechanism with five rules to drive the population toward higher classification accuracy. Compared with seven state-of-the-art multi-objective algorithms across 14 microarray datasets, MOGS-MLPSAE achieves superior performance, with classification accuracy 1.56–8.04 % higher than other algorithms and the lowest gene selection rate (average 1 %, minimum 0.01 %). This study demonstrates MOGS-MLPSAE's effectiveness in balancing classification accuracy and gene selection rate, offering a robust solution for microarray-based cancer classification.}, archive = {J_ASOC}, author = {Min Li and Rutun Cao and Chen Jin and Junke Wang and Shaobo Deng and Xiang Yu}, doi = {10.1016/j.asoc.2025.113794}, journal = {Applied Soft Computing}, month = {12}, pages = {113794}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-level pooling self-adaptive evolutionary multi-objective gene selection algorithm for microarray data classification}, volume = {184}, year = {2025}, }
• (2025). Activity transitions for semi-supervised federated learning in sensor-based human activity recognition. ASOC, 184, 113793. (www)

  @article{ , abstract = {Sensor-based Human Activity Recognition (HAR) is a core component in many real-world applications such as healthcare, fitness tracking, and smart environments. However, training effective HAR models is often constrained by limited labeled data and growing privacy concerns. Federated Learning (FL) offers a privacy-preserving solution by enabling collaborative model training without sharing raw sensor data. To further address the label scarcity challenge, we propose ATCoFed, a Semi-Supervised Federated Learning (SSFL) framework that introduces a novel pseudo-label filtering method based on activity transition patterns. Unlike existing approaches that rely solely on confidence thresholds, ATCoFed incorporates temporal context by using Long Short-Term Memory (LSTM) networks and Large Language Models (LLMs) as data-driven evaluators to validate the consistency of predicted activity sequences. This dual-evaluator mechanism improves the quality of pseudo-labels and enhances model robustness. Experimental results on benchmark HAR datasets demonstrate that ATCoFed consistently outperforms SSFL baselines, achieving better accuracy while maintaining computational and communication efficiency. These findings highlight the potential of activity-aware filtering to improve semi-supervised learning in privacy-preserving HAR applications.}, archive = {J_ASOC}, author = {Tori Andika Bukit and Ericka Pamela Bermudez Pillado and Bernardo Nugroho Yahya and Seok-Lyong Lee}, doi = {10.1016/j.asoc.2025.113793}, journal = {Applied Soft Computing}, month = {12}, pages = {113793}, shortjournal = {Appl. Soft. Comput.}, title = {Activity transitions for semi-supervised federated learning in sensor-based human activity recognition}, volume = {184}, year = {2025}, }
• (2025). TPCMEA: A tri-population evolutionary algorithm with adaptive stage-switching for complex CMOPs. ASOC, 184, 113792. (www)

  @article{ , abstract = {Constrained multi-objective optimization problems (CMOPs) involve the optimization of objective functions along with the satisfaction of constraints, presenting significant challenges in obtaining optimal solutions. To address difficulties such as conflicting objectives, complex constraints, disconnected and narrow feasible regions, this paper proposes a novel tri-population constrained multi-objective evolutionary algorithm (TPCMEA). This algorithm innovatively introduces an adaptive diversity-convergence stage-switching method (DCSSM), based on dual indicators of diversity and convergence, by dynamically monitoring the Euclidean distance distribution of the populations and convergence towards the optimal objective vector, overcoming the issues of premature convergence or resource wastage in traditional algorithms. Moreover, TPCMEA adopts a collaborative framework with three populations, each utilizing different search strategies to effectively solve complex CMOPs. These strategies encompass achieving rapid convergence under constraints, exploring unconstrained fronts to enhance solution diversity, and integrating information from the first two populations through knowledge transfer and compressive sampling strategies. Experimental results demonstrate that TPCMEA exhibits significant competitive advantages across the MW, LIRCMOP, CF benchmark test suites and real-world engineering problems, not only in terms of superior convergence and diversity but also displaying stronger robustness and adaptability in addressing issues related to narrow feasible regions and disconnected fronts.}, archive = {J_ASOC}, author = {Jun Chen and Xiaobo Li and Yuxin Zhao and Zhendi Ma and Zhongmei Han and Yanxia Bao}, doi = {10.1016/j.asoc.2025.113792}, journal = {Applied Soft Computing}, month = {12}, pages = {113792}, shortjournal = {Appl. Soft. Comput.}, title = {TPCMEA: A tri-population evolutionary algorithm with adaptive stage-switching for complex CMOPs}, volume = {184}, year = {2025}, }
• (2025). Multi-objective differential evolution algorithm integrating a directional generation mechanism for multi-objective optimization problems. ASOC, 184, 113791. (www)

  @article{ , abstract = {Multi-objective Evolutionary Algorithms (MOEAs) have gained significant attention due to their effectiveness in solving multi-objective optimization problems. However, when dealing with complex problems, they often face challenges such as low convergence accuracy and poor diversity. To address these issues, we propose a novel multi-objective differential evolution algorithm, MODE-FDGM, which integrates a directional generation mechanism. The key contributions are: (1) A directional-generation method leverages current and past information to rapidly build feasible solutions, boosting both speed and quality in exploring Pareto non-dominated space; (2) An update mechanism that combines crowding distance evaluation, iterating the population and incorporating historical information to enhance diversity and improve the ability to escape local optima; and (3) The introduction of an ecological niche radius concept along with a dual-mutation ecological niche selection evolution strategy, which improves exploration of unexplored spaces and preserves population diversity. Comparative experiments against 7 algorithms, including both classical and contemporary ones, on 24 benchmark functions demonstrate that the proposed algorithm markedly enhances the exploration of Pareto non-dominated solutions, exhibiting superior performance and advanced capabilities.}, archive = {J_ASOC}, author = {Zhuoxuan Yuan and Haibin Ouyang and Steven Li and Essam H. Houssein and Nagwan Abdel Samee}, doi = {10.1016/j.asoc.2025.113791}, journal = {Applied Soft Computing}, month = {12}, pages = {113791}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-objective differential evolution algorithm integrating a directional generation mechanism for multi-objective optimization problems}, volume = {184}, year = {2025}, }
• (2025). Relaxation-based exploration and clustering-based exploitation for multimodal multi-objective optimization. ASOC, 184, 113790. (www)

  @article{ , abstract = {Multimodal multi-objective optimization problems represent a specific type of multi-objective optimization problem where multiple distinct Pareto optimal solution sets (PSs) correspond to the same Pareto optimal front. Although the local PSs may not be as effective as the global PSs, obtaining the latter can be more costly. Therefore, the local PSs still hold some value for decision makers. However, many multimodal multi-objective evolutionary algorithms (MMOEAs) adopt the convergence-first criterion, which makes it challenging to obtain both the global and local PSs. To address the above issue, this paper proposes a two-stage dual-population coevolutionary algorithm with relaxation-based exploration and clustering-based exploitation (RCEA). Specifically, the relaxation-based exploration strategy defines an adaptive maximum Pareto rank, and relaxes the selection threshold for candidates from strictly non-dominated solutions to valuable solutions within this rank. Moreover, the clustering-based exploitation strategy performs a density-based decomposition on the population, separating different global and local PSs to facilitate more targeted evolution. During exploitation, a novel offspring generation operator is employed for intra-subpopulation crossover, preventing the waste of computational resources. Additionally, an improved convergence indicator is proposed to measure the convergence of solutions, which effectively handles the isolated solutions that may be erroneously identified as local Pareto optimal solutions by adaptively adjusting the neighborhood radius. Experimental results demonstrate that RCEA outperforms seven state-of-the-art MMOEAs on 70 benchmark problems.}, archive = {J_ASOC}, author = {Qianlong Dang and Zhiyang Zhang and Xiaochuan Gao and Tingting Wang}, doi = {10.1016/j.asoc.2025.113790}, journal = {Applied Soft Computing}, month = {12}, pages = {113790}, shortjournal = {Appl. Soft. Comput.}, title = {Relaxation-based exploration and clustering-based exploitation for multimodal multi-objective optimization}, volume = {184}, year = {2025}, }
• (2025). Compute-efficient and backpropagation-free pseudoinverse learning for neural networks: A comprehensive survey. ASOC, 184, 113789. (www)

  @article{ , abstract = {The pseudoinverse learning algorithm is a non-gradient, efficient learning scheme originally designed for training single hidden layer feedforward neural networks. It has been developed into various variants and successfully applied across numerous fields. This paper provides a systematic review of the fundamental theories of the pseudoinverse learning algorithm and its major variants, outlining different types of neural networks and learning system architectures based on the pseudoinverse learning scheme. Furthermore, we summarize the fundamental ideas and methodologies of applying the pseudoinverse learning scheme to various learning tasks such as classification, representation learning, time series forecasting, incremental learning, automated machine learning, and content generation. We also summarize and compare the performance of pseudoinverse learning with representative competing baselines on several commonly used data sets based on existing literature reports. The results demonstrate that PIL exhibits significant efficiency advantages over gradient-based approaches (training time was reduced by 72.73% to 99.37%), aligning with its inherent gradient-free nature. Notably, recent PIL variants maintain this computational superiority while achieving enhanced performance compared to other gradient-free algorithms. In addition, we briefly introduce the representative applications of pseudoinverse learning in various fields. To the best of our knowledge, this is the first comprehensive review in this field to encompass all aforementioned aspects. It facilitates the synthesis and integration of existing knowledge from disparate studies. By highlighting limitations in prior works including the computational complexity in large-scale pseudoinverse computation, potential numerical instability for ill-conditioned matrices, risk of overfitting, and constraints in modeling multidimensional patterns, this paper also recommends directions for future research in this area.}, archive = {J_ASOC}, author = {Ke Wang and Pandi Liu and Mohammed A.B. Mahmoud and Ping Guo and Yafei Li}, doi = {10.1016/j.asoc.2025.113789}, journal = {Applied Soft Computing}, month = {12}, pages = {113789}, shortjournal = {Appl. Soft. Comput.}, title = {Compute-efficient and backpropagation-free pseudoinverse learning for neural networks: A comprehensive survey}, volume = {184}, year = {2025}, }
• (2025). An efficient constrained multi-population cooperation search algorithm for multi-objective optimization of multireservoir operation. ASOC, 184, 113788. (www)

  @article{ , abstract = {In multi-objective optimization of multireservoir operation, complex nonlinear constraints often result in highly fragmented and discontinuous feasible regions, making it difficult for traditional algorithms to simultaneously achieve rapid convergence and maintain solution diversity. To effectively address this challenge, this paper proposes a novel Constrained Multi-Population Cooperation Search Algorithm (cMPCSA), which establishes an adaptive balance between global exploration and local exploitation through a cooperative evolutionary mechanism. The proposed algorithm achieves key advancements by constructing a modular multi-population framework with distinct subpopulation roles, integrating a constraint-independent environmental selection mechanism, and introducing a multi-factor learning strategy that combines edge, mean, and stochastic solutions to guide the search process more effectively. Experimental results on four mainstream multi-objective benchmark test suites demonstrate that cMPCSA significantly outperforms existing algorithms in terms of convergence performance and solution quality. Further application to a real-world multireservoir operation case confirms its robustness and superiority in coordinating multiple objectives such as hydropower generation, water supply reliability, and ecological flow protection. Overall, this study provides an efficient and scalable algorithmic tool for addressing complex multi-objective optimization problems in modern water resource management.}, archive = {J_ASOC}, author = {Zhong-kai Feng and Li Zhang and Xia-yu Wang and Fang Yang and Yi-hong Jiang and Sen Wang and Wen-jing Niu}, doi = {10.1016/j.asoc.2025.113788}, journal = {Applied Soft Computing}, month = {12}, pages = {113788}, shortjournal = {Appl. Soft. Comput.}, title = {An efficient constrained multi-population cooperation search algorithm for multi-objective optimization of multireservoir operation}, volume = {184}, year = {2025}, }
• (2025). Deep insights into cognitive decline: A survey of leveraging non-intrusive modalities with deep learning techniques. ASOC, 184, 113787. (www)

  @article{ , abstract = {Cognitive decline is a natural part of aging. However, under some circumstances, this decline is more pronounced than expected, typically due to disorders such as Alzheimer’s disease. Early detection of an anomalous decline is crucial, as it can facilitate timely professional intervention. While medical data can help, it often involves invasive procedures. An alternative approach is to employ non-intrusive techniques such as speech or handwriting analysis, which do not disturb daily activities. This survey reviews the most relevant non-intrusive methodologies that use deep learning techniques to automate the cognitive decline detection task, including audio, text, and visual processing. We discuss the key features and advantages of each modality and methodology, including state-of-the-art approaches like Transformer architecture and foundation models. In addition, we present studies that integrate different modalities to develop multimodal models. We also highlight the most significant datasets and the quantitative results from studies using these resources. From this review, several conclusions emerge. In most cases, text-based approaches consistently outperform other modalities. Furthermore, combining various approaches from individual modalities into a multimodal model consistently enhances performance across nearly all scenarios.}, archive = {J_ASOC}, author = {David Ortiz-Perez and Manuel Benavent-Lledo and Jose Garcia-Rodriguez and David Tomás and M. Flores Vizcaya-Moreno}, doi = {10.1016/j.asoc.2025.113787}, journal = {Applied Soft Computing}, month = {12}, pages = {113787}, shortjournal = {Appl. Soft. Comput.}, title = {Deep insights into cognitive decline: A survey of leveraging non-intrusive modalities with deep learning techniques}, volume = {184}, year = {2025}, }
• (2025). Improving adversarial transferability with neighborhood gradient information. ASOC, 184, 113786. (www)

  @article{ , abstract = {Deep neural networks (DNNs) are known to be susceptible to adversarial examples, leading to significant performance degradation. In black-box attack scenarios, a considerable attack performance gap between the surrogate model and the target model persists. This work focuses on enhancing the transferability of adversarial examples to narrow this performance gap. We observe that the gradient information around the clean image, i.e., Neighborhood Gradient Information (NGI) , can offer high transferability. Based on this insight, we introduce NGI-Attack, incorporating Example Backtracking and Multiplex Mask strategies to exploit this gradient information and enhance transferability. Specifically, we first adopt Example Backtracking to accumulate Neighborhood Gradient Information as the initial momentum term. Then, we utilize Multiplex Mask to form a multi-way attack strategy that forces the network to focus on non-discriminative regions, which can obtain richer gradient information during only a few iterations. Extensive experiments demonstrate that our approach significantly enhances adversarial transferability. Especially, when attacking numerous defense models, we achieve an average attack success rate of 95.2%. Notably, our method can seamlessly integrate with any off-the-shelf algorithm, enhancing their attack performance without incurring extra time costs.}, archive = {J_ASOC}, author = {Haijing Guo and Jiafeng Wang and Zhaoyu Chen and Kaixun Jiang and Lingyi Hong and Pinxue Guo and Jinglun Li and Wenqiang Zhang}, doi = {10.1016/j.asoc.2025.113786}, journal = {Applied Soft Computing}, month = {12}, pages = {113786}, shortjournal = {Appl. Soft. Comput.}, title = {Improving adversarial transferability with neighborhood gradient information}, volume = {184}, year = {2025}, }
• (2025). Federated semi-supervised learning for in-domain/Cross-domain person re-identification. ASOC, 184, 113785. (www)

  @article{ , abstract = {The Person Re-identification (Re-ID) task refers to retrieving images of the same person from different camera views. However, the task faces three major challenges: building a person dataset may infringe on personal privacy, collecting large-scale datasets for identity annotation requires high costs, and low quality local datasets can lead to performance bottlenecks. To address the above challenges, we propose Federated semi-supervised frameworks for the first time to solve in-domain/Cross-domain Re-ID problems, which can reduce annotation costs while protecting privacy. Firstly, we propose two different weighted aggregation federated learning strategies. Specifically, under the in-domain setting, labeled clients are utilized to determine the quality of unlabeled clients. Under the cross-domain setting, the model’s recognition capability is extended to each unlabeled domain by measuring inter-domain differences. In addition, a cascaded Global-personalized Model System and Personalized Model Knowledge Transfer Module are proposed to further address the domain-gap issue by sufficiently decoupling the global shared knowledge and the personalized knowledge unique to each domain. Finally, a Classifier Checking Module is proposed to significantly reduce computational costs by dynamically adjusting the pseudo-labels of unlabeled samples. Extensive experiments on multiple public datasets of varying scales and quality verify that the proposed frameworks can achieve acceptable performance at a lower cost while ensuring the security of personal data. Meanwhile, the effectiveness of the proposed training strategies and modules in other tasks is also verified, which proves their inspirational significance for research and applications in the fields of weakly supervised learning, user privacy protection and intelligent video surveillance.}, archive = {J_ASOC}, author = {Xinyuan Chen and Mingwen Shao and Yi Niu and Qiao Zhang}, doi = {10.1016/j.asoc.2025.113785}, journal = {Applied Soft Computing}, month = {12}, pages = {113785}, shortjournal = {Appl. Soft. Comput.}, title = {Federated semi-supervised learning for in-domain/Cross-domain person re-identification}, volume = {184}, year = {2025}, }
• (2025). Attention-guided triplet framework for synthetic data generation and semantic segmentation of railway fasteners under data scarcity and disparity. ASOC, 184, 113784. (www)

  @article{ , abstract = {Railway infrastructure monitoring faces significant challenges due to limited data availability, labor-intensive annotations, and imbalanced datasets—factors that hinder efficient fastener maintenance. To address these issues, this study proposes a novel three-phase framework that leverages attention-guided deep learning for the generation and analysis of synthetic railway fastener images. In the first phase, an attention-based Deep Convolutional Generative Adversarial Network (DCGAN) is introduced to generate high-fidelity synthetic images that closely mimic real-world conditions. Unlike conventional GANs, the attention mechanism enables the model to focus on critical structural features of fasteners, enhancing the realism and diversity of the generated data. The second phase applies advanced denoising techniques, with the DnCNN model outperforming traditional methods like Median Filtering in preserving fine details. The final phase employs a Convolutional Autoencoder (CAE) for accurate semantic segmentation, achieving 88.8 % accuracy on the synthetic dataset. This end-to-end methodology improves model generalizability, reduces reliance on manual labeling, and provides a cost-effective solution for automated railway inspection. By bridging the gap between real and synthetic data, it also lays the groundwork for scalable, intelligent infrastructure monitoring systems, supporting the advancement of safer and more efficient railway operations.}, archive = {J_ASOC}, author = {Qasim Zaheer and Momina Malik and S.Muhammad Ahmed Hassan Shah and Chengbo Ai and Hongzhi Wang and Zhiyu Liang and Shi Qiu}, doi = {10.1016/j.asoc.2025.113784}, journal = {Applied Soft Computing}, month = {12}, pages = {113784}, shortjournal = {Appl. Soft. Comput.}, title = {Attention-guided triplet framework for synthetic data generation and semantic segmentation of railway fasteners under data scarcity and disparity}, volume = {184}, year = {2025}, }
• (2025). An Actor–Critic-based adapted deep reinforcement learning model for multi-step traffic state prediction. ASOC, 184, 113783. (www)

  @article{ , abstract = {Traffic state prediction is critical to decision-making in various traffic management applications. Despite significant advancements in Deep Learning (DL) models, such as Long Short-Term Memory (LSTM), Graph Neural Networks (GNN), and attention-based transformer models, multi-step predictions remain challenging. The state-of-the-art models face a common limitation: the predictions’ accuracy decreases as the prediction horizon increases, a phenomenon known as error accumulation. In addition, with the arrival of non-recurrent events and external noise, the models fail to maintain good prediction accuracy. Deep Reinforcement Learning (DRL) has been widely applied to diverse tasks, including optimising intersection traffic signal control. However, its potential to address multi-step traffic prediction challenges remains underexplored. This study introduces an Actor–Critic-based adapted DRL method to explore the solution to the challenges associated with multi-step prediction. The Actor network makes predictions by capturing the temporal correlations of the data sequence, and the Critic network optimises the Actor by evaluating the prediction quality using Q-values. This novel combination of Supervised Learning and Reinforcement Learning (RL) paradigms, along with non-autoregressive modelling, helps the model to mitigate the error accumulation problem and increase its robustness to the arrival of non-recurrent events. It also introduces a Denoising Autoencoder to deal with external noise effectively. The proposed model was trained and evaluated on three benchmark traffic flow and speed datasets. Baseline multi-step prediction models were implemented for comparison based on performance metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The results reveal that the proposed method outperforms the baselines by achieving average improvements of 0.26 to 21.29% in terms of MAE and RMSE for up to 24 time steps of prediction length on the three used datasets, at the expense of relatively higher computational costs. On top of that, this adapted DRL approach outperforms traditional DRL models, such as Deep Deterministic Policy Gradient (DDPG), in accuracy and computational efficiency.}, archive = {J_ASOC}, author = {Selim Reza and Marta Campos Ferreira and J.J.M. Machado and João Manuel R.S. Tavares}, doi = {10.1016/j.asoc.2025.113783}, journal = {Applied Soft Computing}, month = {12}, pages = {113783}, shortjournal = {Appl. Soft. Comput.}, title = {An Actor–Critic-based adapted deep reinforcement learning model for multi-step traffic state prediction}, volume = {184}, year = {2025}, }
• (2025). Grammatical evolution for automatic design of actuated traffic signal control plans. ASOC, 184, 113782. (www)

  @article{ , abstract = {In traffic networks, proper signal control design is essential to ensure a reasonable level of service. Signal control designs are becoming increasingly complex, with numerous settings that must be calibrated and set. This paper introduces a novel approach for handling this complexity by automatically generating optimal actuated signal control plans using Grammatical Evolution (GE). GE has proven its effectiveness in automating the design of different complex systems, such as neural networks and analog electronic circuits. GE’s distinctive mapping and representation capabilities make it a powerful candidate for optimizing various systems. In contrast to traditional optimization methods for actuated signal plans, which focus on specific parameters, such as green times and cycle length, the GE-based approach evolves complete plans, including phases, detector placements, and transit priority strategies. As a result, it eliminates the need for human intervention in the design process, making it more efficient and less time-consuming. The proposed approach was tested with an application to an isolated intersection in Haifa, Israel. The results showed that the automatically generated signal plan outperformed the existing plan by reducing delay times and queue lengths. Moreover, this method demonstrated its efficiency in generating reliable traffic signal plans under challenging traffic conditions.}, archive = {J_ASOC}, author = {Mahmud Keblawi and Tomer Toledo}, doi = {10.1016/j.asoc.2025.113782}, journal = {Applied Soft Computing}, month = {12}, pages = {113782}, shortjournal = {Appl. Soft. Comput.}, title = {Grammatical evolution for automatic design of actuated traffic signal control plans}, volume = {184}, year = {2025}, }
• (2025). Optimized resource management in containerized clouds via hierarchical autoregressive network based workload prediction. ASOC, 184, 113781. (www)

  @article{ , abstract = {The swift growth of containerized applications calls for highly precise and real-time prediction of workloads to make resource allocation in cloud environments efficient. Current methods are deficient either by not being able to catch the complex long-term workload dependencies in patterns or by introducing too much computational overhead that degrades real-time performance. We propose a new integrated model where a transformer-based Hierarchical Autoregressive Network (HARN) is dynamically fused with TES. With the use of self-attention mechanisms in modeling complex temporal dynamics and long-range dependencies, on the one hand, HARN captures the detailed seasonal variation but robustly deals with the smoother short-term oscillations, giving TES ample opportunity to function. The model dynamically adjusts their contributions to fit the best weighting in real-time, hence leading to optimal predictability even for changing workloads. Our analysis on the publicly released Alibaba v2018 dataset shows an average improvement in Mean Absolute Error (MAE) of 15.50% and an improvement in Root Mean Square Error (RMSE) of 14.80%. Furthermore, experiments on a specially designed container dataset show even more significant improvements, with improvements of up to 25% in CPU metrics and 40% in memory metrics. These findings and minimal computational overhead emphasize the model’s ability to facilitate real-time proactive resource provisioning in dynamic cloud settings.}, archive = {J_ASOC}, author = {Shivani Tripathi and Angelina Shibu and Priyadarshni and Rajiv Misra and T.N. Singh}, doi = {10.1016/j.asoc.2025.113781}, journal = {Applied Soft Computing}, month = {12}, pages = {113781}, shortjournal = {Appl. Soft. Comput.}, title = {Optimized resource management in containerized clouds via hierarchical autoregressive network based workload prediction}, volume = {184}, year = {2025}, }
• (2025). Pervasive multifaceted process based generative adversarial network for image quality enhancement. ASOC, 184, 113780. (www)

  @article{ , abstract = {The practice of Generative Adversarial Networks (GAN) has extended a lot of consideration in recent times. In most of the GAN methods are problem-specific related that are personalized to report numerous trials of individual application rather than performing other image improvement tasks. Furthermore, the basic GAN generators influence their boundaries in numerous image restoration and development use cases. Therefore, in this paper, we propose a generic GAN referred to as Pervasive Multifaceted Process based Generative Adversarial Network (PMPGAN). In this generator, we introduced multiple Convolutional Neural Networks (CNN) followed by Multi-dimensional Pyramid Pooling Module (MPPM) and Attention Module (AM), of which the input for the AM is given as low-level features and it produces output with enhanced feature map. Meanwhile, we enhanced the improvement outcome of the generated image with discriminator loss function. Finally, we tested the efficiency of the proposed system through extensive experiments on five challenging applications for image enhancement, image restoration, and infrared image translation to determine the dominance and efficiency in eliminating image degradation and producing visually interesting fake images. Our PMPGAN quantitatively outperforms several latest models. The results show that the PMPGAN model is superior to the existing models.}, archive = {J_ASOC}, author = {K. Balaji}, doi = {10.1016/j.asoc.2025.113780}, journal = {Applied Soft Computing}, month = {12}, pages = {113780}, shortjournal = {Appl. Soft. Comput.}, title = {Pervasive multifaceted process based generative adversarial network for image quality enhancement}, volume = {184}, year = {2025}, }
• (2025). Long-term time series forecasting by a frequency-domain enhanced temporal convolutional network with the stationary residual regularization. ASOC, 184, 113779. (www)

  @article{ , abstract = {The current time series models can resolve the long-term time series forecasting problem by leveraging artificial neural networks (ANNs) to capture complex temporal dependencies. However, the state-of-the-art ANNs for long-term prediction, Transformer-based models, have two inherent shortcomings that undermine both credibility and accuracy: losing temporal order information and ignoring information in prediction residuals. To fill these gaps, this study proposes a frequency-domain enhanced multi-scale temporal convolutional network (FMTCN) and a hybrid loss function regularized by stationary residuals called QM. Firstly, to preserve temporal order information, a multi-scale block with dilated causal convolution is developed as the core component of FMTCN, which can ensure the dependencies of predictions on sequentially ordered data. Secondly, an adaptive fusion mechanism for time and frequency domain patterns is designed to bridge the performance gap between convolution-based and Transformer-based models. Thirdly, to reduce residual autocorrelation, the proposed hybrid loss function integrates the Ljung–Box statistic into the mean squared error loss function as the regularization. Finally, extensive experiments across five real-world datasets are conducted to validate the proposed methods. Compared with the state-of-the-art method, the proposal improves the accuracy by 8.6% in the univariate long-term prediction and 6.1% in the multivariate long-term prediction.}, archive = {J_ASOC}, author = {Jing Wang and Yanbing Ju and Peiwu Dong and Tian Ju}, doi = {10.1016/j.asoc.2025.113779}, journal = {Applied Soft Computing}, month = {12}, pages = {113779}, shortjournal = {Appl. Soft. Comput.}, title = {Long-term time series forecasting by a frequency-domain enhanced temporal convolutional network with the stationary residual regularization}, volume = {184}, year = {2025}, }
• (2025). Loss functions in classification: An comprehensive overview and comparative study. ASOC, 184, 113778. (www)

  @article{ , abstract = {Loss functions are among the most fundamental components of a classifier’s learning processes and consequently can significantly affect the performance of classification methods. However, despite the substantial potential impact of loss functions on the classification accuracy, they have been incomprehensively investigated in the literature. For this reason, this paper tries to comprehensively evaluate the potency and impact of loss functions on the performance and accuracy in classification problems. For this purpose, fifty-five different loss functions in twelve categories—Linear Continuous Distance, Nonlinear Continuous Distance, Linear Semi-Continuous Distance, Nonlinear Semi-Continuous Distance, Linear Discrete Distance, Nonlinear Discrete Distance, Linear Continuous Direction, Nonlinear Continuous Direction, Linear Semi-Continuous Direction, Nonlinear Semi-Continuous Direction, Linear Discrete Direction, and Nonlinear Discrete Direction, are considered. In addition, in this paper, six distinct environmental pollutants/pollution benchmark data sets, three classifier types—statistical, shallow intelligence, and deep intelligence, are exemplary considered. Furthermore, in this paper, the most important measure of regular classification problems, the classification rate, has been chosen in order to compare these loss functions. The experimental results have unequivocally validated that the choice of loss functions has a significant impact on accuracy and classification rates. Numerical results of loss functions indicate that the difference between the lowest and the highest classification rate in the statistical, shallow intelligent, and deep learning classifiers, is averagely equal to 6.41 %, 5.16 %, and 2.24 %, respectively. It means that the model designer, by choosing the appropriate loss function, can averagely improve more than 4 % the obtained classification rate. Empirical results show that in the general perspective, the lowest performances of loss functions are overall related to the linear, continuous, and distance-based categories, in contrast to the highest ones, which are nonlinear, discrete, and direction-based. The exception to this outcome is the Zero-One family, in which all their loss functions, i.e., linear/nonlinear, distance/direction, are equivalent and yield the same accuracy. The best classification rate is also related to this family, that significantly better than other loss functions in all cases, as well as all classifiers. The Zero-One loss functions can averagely achieve a 98.62 % classification rate that 17.65 % is averagely higher than others. These evidences illustrate that, in addition to other effective features and factors, the loss function type should also be considered by environmental model designers in order to yield a more desired classification rate.}, archive = {J_ASOC}, author = {Fatemeh Chahkoutahi and Mehdi Khashei and Naser Molaverdi}, doi = {10.1016/j.asoc.2025.113778}, journal = {Applied Soft Computing}, month = {12}, pages = {113778}, shortjournal = {Appl. Soft. Comput.}, title = {Loss functions in classification: An comprehensive overview and comparative study}, volume = {184}, year = {2025}, }
• (2025). Ultra-low memory spatiotemporal decomposition recurrent neural networks for edge structural fault monitoring. ASOC, 184, 113777. (www)

  @article{ , abstract = {The hardware memory resources of microcontrollers in wireless sensor network nodes are limited, currently only capable of data acquisition and simple computing, making it difficult to perform neural network inference for edge fault monitoring. Running neural networks on microcontrollers can enhance the edge data processing capabilities of the nodes. Recurrent neural networks with short and medium term memory excel at processing sequential data. However, inference on microcontrollers consumes a significant amount of memory, necessitating solutions to the memory constraints. This study proposes a spatiotemporal decomposition method for recurrent neural networks to address the memory constraint issue when performing inference on resource-constrained nodes, thus enabling edge fault monitoring. The method decomposes recurrent neural networks in spatiotemporal dimensions, significantly reducing memory usage during operation. Additionally, it proposes model parameter storage and addressing techniques to ensure accurate reading of Flash data. Experiments have verified that this method can achieve 99.7 % accuracy in edge fault classification, consuming only 332 bytes of RAM and 768 bytes of Flash. The spatiotemporal decomposition method also provides a solution for running other time-series models on resource-constrained edge devices.}, archive = {J_ASOC}, author = {Hao Fu and Lei Deng and Baoping Tang and Shuaiwen Cui and Yuguang Fu}, doi = {10.1016/j.asoc.2025.113777}, journal = {Applied Soft Computing}, month = {12}, pages = {113777}, shortjournal = {Appl. Soft. Comput.}, title = {Ultra-low memory spatiotemporal decomposition recurrent neural networks for edge structural fault monitoring}, volume = {184}, year = {2025}, }
• (2025). Hybrid SpinalNet-fuzzy-shufflenet for brain tumor detection using MRI images. ASOC, 184, 113775. (www)

  @article{ , abstract = {Today, a significant portion of the global population is affected by the serious medical condition known as Brain Tumor (BT). Damage to healthy brain tissue is suspected, as it is currently the most important cause of a huge quantity of mortality. To prevent patients from dying, early detection is very essential. Despite various notable efforts and hopeful results in this area, accurate segmentation and categorization remain a difficult task. To address these gaps, a SpinalNet Fuzzy Shufflenet (SFShuffleNet) is proposed for the detection of BT. First, the input image is fed into the preprocessing stage, utilizing ROI extraction. Subsequently, the preprocessed image undergoes enhancement using histogram equalization techniques. Then, the enhanced image undergoes segmentation with Fuzzy Local Information C-Means (FLICM). Following segmentation, the image is augmented through sharpening, translation, random erasing, and resizing techniques. Features such as mean, variance, standard deviation, average, contrast, skewness, kurtosis, and entropy with the Local Frequency Descriptor (LFD) are extracted. Finally, the proposed SFShuffleNet, which combines SpinalNet and Shufflenet with fuzzy concept modifications, detects BT. SFShuffleNet achieved the highest accuracy, sensitivity, specificity, and F1-score of 91.46 %, 90.96 %, 92.78 %, and 90.79 %, respectively.}, archive = {J_ASOC}, author = {P Srinivasa Rao and Swathi Sowmya Bavirthi and G. Sharada and Ponnaboyina Ranganath and Vemuri Sailaja and G Vimala Kumari}, doi = {10.1016/j.asoc.2025.113775}, journal = {Applied Soft Computing}, month = {12}, pages = {113775}, shortjournal = {Appl. Soft. Comput.}, title = {Hybrid SpinalNet-fuzzy-shufflenet for brain tumor detection using MRI images}, volume = {184}, year = {2025}, }
• (2025). Enzyme classification integrating LSTM and prot-BERT sequence encoding. ASOC, 184, 113774. (www)

  @article{ , abstract = {Enzyme classification is essential for deciphering cellular and biological processes, driving targeted research, and influencing domains such as drug discovery and bioengineering. While various automated tools exist for enzyme classification, many are limited in scope or no longer operational. This study utilizes advanced artificial intelligence (AI) algorithms, including SVM, RF, simpleRNN, LSTM, ConvLSTM, and bidirectional LSTM, combined with numeric and Prot-BERT-based protein sequence encodings, to classify 1991 enzymes across 7 main classes, 69 subclasses, 216 sub-subclasses, and 1333 substrates. Among all trained models, the bidirectional LSTM model integrated with Prot-BERT sequence encoding (ECiLPSE) demonstrated exceptional accuracy of 99.14 % on the training set and 98.41 % on the test set, effectively capturing intricate sequence details. In addition to high accuracy, ECiLPSE achieved an F1-score of 0.99 (training set) and 0.98 (test set), with AU-ROC scores of 0.98 and 0.97, respectively. Precision and recall were both 0.99 on the training set and 0.98 on the test set. The 95 % confidence interval for test set accuracy (98.00 % - 99.97 %) further supports the model’s robustness and discriminative capability across classes. To benchmark performance, ECiLPSE was evaluated against four EC prediction tools (ECPred, CLEAN, EZYPred, EzyDeep), demonstrating superior accuracy and computational efficiency. The results from three case studies (using New-1277, Price-149, and halogenases datasets) highlight the limitations of the existing tools, underscoring the predictive accuracy, reliability, and applicability of ECiLPSE. ECiLPSE is available as a standalone tool for large datasets and as a web server for small datasets, offering a robust resource for enzyme classification, advancing computational approaches in enzyme research, and promising applications in drug discovery.}, archive = {J_ASOC}, author = {Anju Sharma and Vineet Diwakar and Rajnish Kumar and Prabha Garg}, doi = {10.1016/j.asoc.2025.113774}, journal = {Applied Soft Computing}, month = {12}, pages = {113774}, shortjournal = {Appl. Soft. Comput.}, title = {Enzyme classification integrating LSTM and prot-BERT sequence encoding}, volume = {184}, year = {2025}, }
• (2025). Intelligent urban GNSS measurement uncertainty prediction by exploring the spatial characteristics with transformer. ASOC, 184, 113773. (www)

  @article{ , abstract = {Global navigation satellite system (GNSS) positioning accuracy is essential for civil applications. Unfortunately, in urban areas, GNSS signals are easily blocked and reflected by tall buildings, namely non-line-of-sight (NLOS) receptions. Thus, it is essential to classify these contaminated receptions and mitigate their pseudorange error for positioning improvement. This paper designed a Transformer-based network that utilizes satellite spatial characteristics to predict GNSS measurement uncertainty. The proposed method accurately classifies 89 % of satellites’ visibility and around 45 % compensation in NLOS pseudorange error, outperforming the state-of-the-art algorithms. By analyzing the attention matrix generated by Transformer, we explore how Transformer utilizes spatial characteristics for satellite measurement uncertainty prediction, which regards satellite visibility and pseudorange as different tasks for training.}, archive = {J_ASOC}, author = {Zekun Zhang and Penghui Xu and Guohao Zhang}, doi = {10.1016/j.asoc.2025.113773}, journal = {Applied Soft Computing}, month = {12}, pages = {113773}, shortjournal = {Appl. Soft. Comput.}, title = {Intelligent urban GNSS measurement uncertainty prediction by exploring the spatial characteristics with transformer}, volume = {184}, year = {2025}, }
• (2025). Modeling self-supervised learning for handling the concept drift with ternary-adaptive ensemble. ASOC, 184, 113772. (www)

  @article{ , abstract = {Over the last several decades, drift detection and adaptation models have been trained on historical drift labels struggle with real-world data streams due to evolving distributions, impacting accurate future predictions. Existing methods update predictive models but often fail to retain previously learned patterns from streaming data, hindered by insufficient predefined labels for categorizing drift detection. To address these constraints, the proposed work introduces a two-task framework comprising a Self-Supervised Learning (SSL) model with a pretext task and a downstream task to efficiently adapt the model to evolving data distributions. In the pretext task, the proposed approach learns patterns from the unlabeled offline data in a self-supervised manner and utilizes the knowledge from the performance-based drift detector. In a subsequent downstream task, the model that has been enhanced with a forgetting-aware Ternary-Adaptive Ensemble (TAE) learner without requiring the drift labels. In the proposed approach, the TAE algorithm comprises drift-aware model updation methods for the Long Short-Term Memory (LSTM) to resolve forgetting in sequential data streams, and the three updation methods involve the pretext model’s weight transfer, Bayesian Optimization-based hyperparameter tuning, and Elastic Weight Consolidation (EWC). The proposed approach effectively handles drifts and outperformed while testing on the real-world weather and synthetic, mixed drift datasets with 82.6% and 92.36% accuracy, respectively.}, archive = {J_ASOC}, author = {Shubhangi Suryawanshi and Anurag Goswami and Pramod Patil}, doi = {10.1016/j.asoc.2025.113772}, journal = {Applied Soft Computing}, month = {12}, pages = {113772}, shortjournal = {Appl. Soft. Comput.}, title = {Modeling self-supervised learning for handling the concept drift with ternary-adaptive ensemble}, volume = {184}, year = {2025}, }
• (2025). Cross-domain knowledge distillation for domain adaptation with GCN-driven MLP generalization. ASOC, 184, 113771. (www)

  @article{ , abstract = {Knowledge distillation (KD) and domain adaptation (DA) represent potential research directions for reducing costs associated with deploying deep neural networks (DNN) in real-world applications. KD focuses on model compression, exploring methods to transfer informative representations from a complex model to a lighter one without incurring additional costs. Conversely, DA emphasizes the data distribution perspective, aiming to decrease labeling expenses by leveraging knowledge extracted from a labeled source domain to minimize classification errors in an unlabeled target domain. In this paper, we introduce a novel knowledge distillation (KD) approach with a teacher–student paradigm for domain adaptation (DA) tasks, termed Improved Cross-domain Knowledge Distillation (ICDKD). Specifically, we employ a graph convolutional network (GCN) classifier as the teacher model and a multilayer perceptron (MLP) classifier as the student model. During training, the teacher model utilizes a message-passing mechanism to capture the topology of the training data through neighbor information, thus explicitly enhancing semantic representations in each category to improve classification accuracy. Subsequently, the extracted knowledge from the GCN teacher model is distilled to the MLP student model. Finally, in the inference stage, only the MLP student model is utilized to meet the latency constraints of applications. Our proposed method effectively combines the strengths of both GCN and MLP classifiers to improve the classification performance and satisfy the real-world application requirements. We implemented our method on various DA benchmark datasets under unsupervised and semi-supervised domain adaptation settings, including ImageCLEF-DA, Office-31, Office-Home, VisDA2017, and DomainNet. The experimental results demonstrate the effectiveness of our proposed method on both CNN-based and ViT-based architectures, achieving outstanding classification performance compared to prior state-of-the-art domain adaptation methods.}, archive = {J_ASOC}, author = {Ba Hung Ngo and Tae Jong Choi}, doi = {10.1016/j.asoc.2025.113771}, journal = {Applied Soft Computing}, month = {12}, pages = {113771}, shortjournal = {Appl. Soft. Comput.}, title = {Cross-domain knowledge distillation for domain adaptation with GCN-driven MLP generalization}, volume = {184}, year = {2025}, }
• (2025). A fuzzy delphi-SERVQUAL model using degree of belief structure for assessing customer satisfaction in automotive after-sales services. ASOC, 184, 113770. (www)

  @article{ , abstract = {This study presents a novel two-phase fuzzy multi-attribute decision-making (MADM) framework for assessing customer satisfaction in automotive after-sales services under uncertainty. The proposed model integrates a modified Fuzzy Delphi Method (FDM) with an enhanced fuzzy SERVQUAL approach, both embedded within a Degree of Belief (DoB) structure to capture evaluative ambiguity and confidence levels in expert and customer judgments. In the first phase, a belief-based FDM enables the efficient screening and prioritization of 58 service quality criteria using a single-round process that incorporates both optimistic and pessimistic expert viewpoints. In the second phase, a belief-driven SERVQUAL model evaluates customer perceptions and expectations from both stringent and lenient perspectives across six dimensions, including a newly introduced Digital Technology dimension that reflects emerging service delivery mechanisms. The model also features a δ-based sensitivity analysis to examine the impact of decision-making attitudes and a seven-zone classification system to categorize service quality gaps with high diagnostic precision. Application of the proposed framework in the Iranian automotive after-sales sector enabled the prioritization of 29 key service quality criteria from an initial pool of 58 indicators, with the most critical factors identified in the ‘Reliability’ and ‘Responsiveness’ dimensions. The resulting insights support evidence-based managerial actions, including the reallocation of resources toward high-impact service areas, targeted digital transformation initiatives, and the formulation of differentiated improvement strategies based on gap severity and belief-based customer expectations .}, archive = {J_ASOC}, author = {Mojtaba Elahi and Ramin Enayati and Mehdi Keramatpour}, doi = {10.1016/j.asoc.2025.113770}, journal = {Applied Soft Computing}, month = {12}, pages = {113770}, shortjournal = {Appl. Soft. Comput.}, title = {A fuzzy delphi-SERVQUAL model using degree of belief structure for assessing customer satisfaction in automotive after-sales services}, volume = {184}, year = {2025}, }
• (2025). Developing a hybrid circular intuitionistic fuzzy framework to assess tourism 5.0 challenges. ASOC, 184, 113769. (www)

  @article{ , abstract = {The adoption of Industry 5.0 (I5.0) in tourism is still in its early stages, and its complex implementation presents challenges that are not fully understood. Existing studies often focus on context-specific challenges, limiting the exploration of broader obstacles. This research aims to identify a comprehensive set of challenges, rank them, and evaluate their interdependencies in I5.0 adoption within tourism. A systematic literature review was conducted to identify key challenges, and expert input was gathered to validate and refine the findings. To evaluate the significance of the identified challenges, a novel extension of the Stepwise Weight Assessment Ratio Analysis (SWARA) method into the Circular Intuitionistic Fuzzy (CIF) environment, referred to as CIF-SWARA, was formulated to determine the relative weights of each challenge under uncertainty with more precision. Additionally, the Decision-Making Trial and Evaluation Laboratory (DEMATEL) method was extended using CIF sets, termed CIF-DEMATEL, to evaluate the causal relationships and relative influence among the challenges. The results from both methods were combined to derive final weights. Furthermore, a conceptual model was constructed using Total Interpretive Structural Modeling (TISM) based on the CIF-DEMATEL output, providing a detailed qualitative assessment of the interactions among the challenges. Findings highlight that “infrastructure issues” are the most urgent challenge requiring attention. Additionally, “technical and technological issues,” “acceptance and adaptability issues,” and “financial issues” emerged as the three most significant challenges for transitioning to “Tourism 5.0.” The model’s robustness was confirmed via sensitivity analysis, and comparative evaluations demonstrated consistent performance against established fuzzy environments. This study advances I5.0 research in tourism by identifying and modeling key challenges, offering valuable insights for policymakers.}, archive = {J_ASOC}, author = {Vahideh Shahin and Dragan Pamucar and Moslem Alimohammadlou}, doi = {10.1016/j.asoc.2025.113769}, journal = {Applied Soft Computing}, month = {12}, pages = {113769}, shortjournal = {Appl. Soft. Comput.}, title = {Developing a hybrid circular intuitionistic fuzzy framework to assess tourism 5.0 challenges}, volume = {184}, year = {2025}, }
• (2025). A self-adaptive discrete artificial bee colony algorithm based on block swap for steelmaking and continuous casting scheduling problem. ASOC, 184, 113768. (www)

  @article{ , abstract = {Due to the high temperature characteristics of the steelmaking continuous casting production, the temperature decreasing energy consumption (TDEC) in the processing process cannot be ignored. Therefore, this paper establishes a mixed integer mathematical model for the steelmaking continuous casting scheduling problem with TDEC (SCCSP TDEC ). The SCCSP TDEC chooses to minimize the objective functions of the total weighted earliness and tardiness, the TDEC and the makespan. To solve the SCCSP TDEC , an improved heuristic based on problem-specific features is proposed in this paper, which takes into account the completion time of the third stage’s cast, and the effect of the cast’s holistic nature. Then, a self-adaptive discrete artificial bee colony algorithm based on block swap (SDABC bs ) is proposed to solve the SCCSP TDEC . This strategy can adaptively select an optimal structure suitable for the current population, and effectively expand the search space of neighborhood solutions. Finally, this paper tests the SDABC bs algorithm based on the extensive instances generated according to the actual production process, and the performance of the SDABC bs algorithm is verified by comparing with other efficient algorithms. The experimental results show that the proposed SDABC bs algorithm is more effective for the SCCSP TDEC among all the algorithms in comparison.}, archive = {J_ASOC}, author = {Yang Yu and Guo-Dong Yang and Qichun Zhang and Liangliang Sun and Xinfu Pang and Yefeng Liu}, doi = {10.1016/j.asoc.2025.113768}, journal = {Applied Soft Computing}, month = {12}, pages = {113768}, shortjournal = {Appl. Soft. Comput.}, title = {A self-adaptive discrete artificial bee colony algorithm based on block swap for steelmaking and continuous casting scheduling problem}, volume = {184}, year = {2025}, }
• (2025). Fine-grained mutation operators for community hiding using genetic algorithms. ASOC, 184, 113767. (www)

  @article{ , abstract = {Community detection plays a key role in uncovering group structures in networks, but its misuse can lead to privacy risks by exposing sensitive relationships. As a proactive defense, community hiding seeks to perturb the network structure to reduce the effectiveness of community detection algorithms. Given the NP-hard nature of this task, genetic algorithms (GAs) widely used due to their robust global search capabilities. However, existing methods lack effective guidance when dealing with vast solution spaces, resulting in inefficient exploration and suboptimal obfuscation outcomes. To address this, we propose Network- T opology- C ombined Community Information H iding A lgorithm (TCHA), a novel GA-based method that leverages node similarity information via node embedding to guide perturbations more effectively. TCHA introduces a multidimensional mutation operator that combines coarse-grained and fine-grained mutation strategies. These fine-grained mutations are performed in the embedding space and decoded back to graph edits, enabling more precise and topologically-aware perturbations. To evaluate the efficiency of this process, we introduce a novel metric based on expected path length within a mutation transition graph, offering deeper insight into evolutionary search dynamics. Experiments on six real-world networks demonstrate that TCHA achieves an average modularity reduction of 31.92% and an average normalized mutual information of 0.6447, outperforming baselines such as Q-Attack, NEURAL, and DICE. These results confirm the superiority of the embedding-guided fine-grained mutation strategy in enhancing community hiding effectiveness.}, archive = {J_ASOC}, author = {Shanqing Yu and Jintao Zhou and Meng Zhou and Yidan Song and Jiaxiang Li and Zeyu Wang and Qi Xuan and Silu Mu and Xiaolei Qian}, doi = {10.1016/j.asoc.2025.113767}, journal = {Applied Soft Computing}, month = {12}, pages = {113767}, shortjournal = {Appl. Soft. Comput.}, title = {Fine-grained mutation operators for community hiding using genetic algorithms}, volume = {184}, year = {2025}, }
• (2025). Predicting scour depth in the presence of aprons using XGBoost-optuna. ASOC, 184, 113766. (www)

  @article{ , abstract = {This paper explores the impact of an apron in mitigating scour downstream of a trapezoidal PK weir through an experimental study. Three apron lengths were tested with two sediment types under varying hydraulic conditions to address local scouring. Results show that longer aprons reduce scouring, especially at lower densimetric Froude numbers, affecting the location of the maximum scour depth and its volume. On average, apron lengths of 1 P , 1.5 P , and 2 P ( P is weir height) decrease the scour hole areas and volumes by approximately 69–77 %. Scour indices decrease by 73–90 % for corresponding apron lengths. New empirical equations have been proposed to aid in apron design, and the estimation of various scour hole geometries . Bayesian Optimized Neural Network (BONN), Extreme Gradient Boosting model tuned by Optuna algorithm (XGBoost-Optuna), and Random Forest were also developed for forecasting scour hole characteristics in the presence of apron. Various regression tests, including residual plots and uncertainty quantification, were imposed to compare the models. The results demonstrated that the XGBoost-Optuna model outperformed the other models, achieving a correlation coefficient ranging from 0.924 to 0.985, a root mean squared error between 0.055 and 5.072, and a mean relative percentage error of 7.14–11.71 %. Most forecasts generated by the XGBoost-Optuna model fell within ±20 % error margins, highlighting its superiority in predicting scour hole characteristics in the presence of the apron for PK weirs.}, archive = {J_ASOC}, author = {Chonoor Abdi Chooplou and Saeed Balahang and Masoud Ghodsian and Mohammad Vaghefi}, doi = {10.1016/j.asoc.2025.113766}, journal = {Applied Soft Computing}, month = {12}, pages = {113766}, shortjournal = {Appl. Soft. Comput.}, title = {Predicting scour depth in the presence of aprons using XGBoost-optuna}, volume = {184}, year = {2025}, }
• (2025). LanPPT: Enhancing landslide crack detection through pyramid pooling transformers. ASOC, 184, 113765. (www)

  @article{ , abstract = {Landslide cracks, also known as tension cracks, are a major part of landslides. Based on the distribution of landslide fractures, the location of a landslide can be broadly described, and the stress distribution of the sliding mass can be inferred. Cracks at a landslide’s head, which are a key indicator of the displacement of the landslide body, can provide early warning signs for landslide hazards. The complete knowledge of the crack development features is still lacking because of many influencing elements and intricate reasons for fracture creation. Even though some early interventions reported models for automated crack identification utilizing advanced machine learning techniques, the problem still has not been solved to its full potential. Thus, an effective deep architecture for landslide crack segmentation is suggested to address these issues, utilizing a synergistic blend of vision transformers and the pyramid pooling concept. In this work, we use the universal vision transformer backbone called the Pyramid Pooling Transformer, and plug it into our pooling-based multi-head spatial attention to build a deep architecture that identifies and segments the landslide cracks, namely LanPPT. Experiments revealed that when a pyramid pooling transformer is used as the backbone network, it performs significantly better than many earlier convolutional neural networks and normal transformer-based networks in various vision tasks. Systematic experiments show that the proposed model achieved superior performance in terms of mIoU and FPS when compared with the chosen state-of-the-art baselines in the landslide crack detection task.}, archive = {J_ASOC}, author = {S. Sreelakshmi and S.S. Vinod Chandra}, doi = {10.1016/j.asoc.2025.113765}, journal = {Applied Soft Computing}, month = {12}, pages = {113765}, shortjournal = {Appl. Soft. Comput.}, title = {LanPPT: Enhancing landslide crack detection through pyramid pooling transformers}, volume = {184}, year = {2025}, }
• (2025). Differential evolution algorithm with cosine similarity-based individual reduction and symmetric uncertainty-based attribute recovery for feature selection. ASOC, 184, 113764. (www)

  @article{ , abstract = {In recent years, with the increase in data scaling, effectively handling high-dimensional datasets has become a focal point of attention. Feature selection (FS), a method for dealing with datasets containing features, has emerged as a crucial technique in fields such as machine learning and data mining with the objective of selecting features that contain richer information while eliminating redundancy. Due to their remarkable performance in global search, evolutionary computation techniques hold substantial potential in the application of FS. However, many existing FS methods overlook the relationships between features. In the context of classification problems, this study presents a novel wrapper FS algorithm based on the differential evolution algorithm. The proposed method reduces redundant features among individuals based on cosine similarity and selectively and recovers certain features in the crossover phase according to the uncertainty similarity. In addition, a probabilistic-based initialization method is designed. The proposed algorithm significantly outperforms five other algorithms in terms of classification error rates over 18 experimental datasets. The experimental results demonstrate a significant enhancement in the performance of the proposed algorithm attributed to these two components.}, archive = {J_ASOC}, author = {Chunzhi Hou and Ziqian Wang and Yu Zhang and Yuki Todo and Jun Tang and Zhenyu Lei and Shangce Gao}, doi = {10.1016/j.asoc.2025.113764}, journal = {Applied Soft Computing}, month = {12}, pages = {113764}, shortjournal = {Appl. Soft. Comput.}, title = {Differential evolution algorithm with cosine similarity-based individual reduction and symmetric uncertainty-based attribute recovery for feature selection}, volume = {184}, year = {2025}, }
• (2025). Diffusion model based on stochastic differential equation with transformer for stock price prediction. ASOC, 184, 113763. (www)

  @article{ , abstract = {The prediction of stock prices is an exemplary interdisciplinary problem, straddling the domains of finance, computer science, econometrics, and mathematics. The fundamental characteristics of stock price data, notably its non-linearity, non-stationarity, and considerable complexity, make the prediction of stock prices an exceptionally challenging task. In recent years, the domain of deep neural networks has shown substantial promise in terms of learning capacities, leading to significant advancements in the field of stock price prediction. However, most existing methods are still limited to predicting the closing price of the next day, rather than the future trend of stock prices, leaving investors with insufficient information for trading decisions. We propose an innovative model, DiffVT, which incorporates a Volatility Transformer for feature extraction from historical data. We introduced a de-stationary attention mechanism that integrates non-stationary information into the model to capture the dependencies in highly volatile stock price sequences. The output is then fed into an improved diffusion model based on stochastic differential equation (SDE). By iteratively solving the reverse-time SDE, our model generates a probabilistic distribution. This approach not only predicts the single-point stock price for the next day but also forecasts the future trend of stock prices over a period, providing a possible range of stock price movements. To our knowledge, DiffVT is the first model to combine a diffusion model with Transformer for stock price prediction. Extensive experiments on multiple stock indices and individual stock datasets demonstrate that DiffVT significantly outperforms state-of-the-art baseline methods, exhibiting excellent performance across various prediction window lengths. Specifically, compared to the second-best models in each domain, our approach achieves reductions of up to 11.73% in Mean Absolute Error (MAE) for point predictions, 17.23% in Continuous Ranked Probability Score (CRPS) for probabilistic predictions, and an improvement of 12.93% in the Sharpe ratio, clearly establishing its superior performance. Our code is available at https://github.com/SoraKsgn/DiffVT/ .}, archive = {J_ASOC}, author = {Chaoyang Wang and Guangyu Liu and Ling Zhu}, doi = {10.1016/j.asoc.2025.113763}, journal = {Applied Soft Computing}, month = {12}, pages = {113763}, shortjournal = {Appl. Soft. Comput.}, title = {Diffusion model based on stochastic differential equation with transformer for stock price prediction}, volume = {184}, year = {2025}, }
• (2025). Multiobjective coevolutionary bayesian learning for hyperspectral sparse unmixing. ASOC, 184, 113762. (www)

  @article{ , abstract = {The capability of multiobjective evolutionary algorithms (MOEAs) to directly address ℓ 0 norm minimization has introduced innovative perspectives for solving sparse problems. However, enhancing the search efficiency of MOEAs remains a formidable challenge, especially in high-dimensional sparse problems. To alleviate the above problem, we propose a novel multiobjective coevolutionary Bayesian learning framework for a classical sparse problem—sparse unmixing. Leveraging the spatial similarity inherent in hyperspectral image patches, the proposed framework alleviates the complexity of the sparse unmixing task by segmenting the original image into homogeneous regions using superpixel segmentation. These regions are then demixed independently and jointly optimized under a cooperative evolutionary paradigm. During the optimization, the row-sparsity parameter inferred through cooperative Bayesian learning is embedded into a specially designed image-level genetic strategy. This parameter can guide the evolutionary direction of the population, encourages the solution to conform to the structural characteristic and significantly improves search efficiency. Experimental results on synthetic and real datasets demonstrated the effectiveness of the proposed algorithm as compared with several sparse unmixing methods.}, archive = {J_ASOC}, author = {Yiting Liu and Maoguo Gong and Xiangming Jiang and Jianzhao Li and Yue Zhao and Yan Pu and Ziqi Di}, doi = {10.1016/j.asoc.2025.113762}, journal = {Applied Soft Computing}, month = {12}, pages = {113762}, shortjournal = {Appl. Soft. Comput.}, title = {Multiobjective coevolutionary bayesian learning for hyperspectral sparse unmixing}, volume = {184}, year = {2025}, }
• (2025). Real-time anomaly detection in seasonal time series with conditional variational autoencoder. ASOC, 184, 113761. (www)

  @article{ , abstract = {Real-time anomaly detection in high-frequency seasonal time series is commonly addressed using prediction-based methods, which require waiting for new values to perform subsequent predictions and demand continuous processing over time. This work introduces a novel framework for real-time anomaly detection in seasonal time series, with a practical implementation using Conditional Variational Autoencoders based on Multilayer Perceptrons. Our approach eliminates the need for historical time series data at inference time, instead generating a one-shot long-term expected time series that enables immediate evaluation of streaming data with minimal computational resources. Empirical evaluations on real-world seasonal time series demonstrate that the proposed approach achieves state-of-the-art performance compared in both semi-supervised and unsupervised settings. The framework provides computational efficiency and low energy consumption, making it suitable for deployment in commodity hardware and offline environments.}, archive = {J_ASOC}, author = {Lorenzo Porcelli and Marcello Trovati and Francesco Palmieri}, doi = {10.1016/j.asoc.2025.113761}, journal = {Applied Soft Computing}, month = {12}, pages = {113761}, shortjournal = {Appl. Soft. Comput.}, title = {Real-time anomaly detection in seasonal time series with conditional variational autoencoder}, volume = {184}, year = {2025}, }
• (2025). Weight optimized stacked LSTM with conditional random fields using self-adaptive generalized normal distribution optimizer for crop yield forecasting. ASOC, 184, 113760. (www)

  @article{ , abstract = {In the agricultural sector, estimating crop production is a difficult task. A crucial element in recent years has forecasted the crop prediction, which is dependent on outside variables include soil, water, and agricultural characteristics. Crop feature extraction is used to predict crop yield using deep learning-based approaches. The predictive ability of the method heavily relies on the nature of the collected information because there is a nonlinear translation between the unprocessed information and crop yield data. Agricultural marketing, crop production, efficient harvest management, and effective fertilization management rely heavily on crop yield forecasts. Numerous manual analyses are used for remote sensing, which is often used for crop prediction. Here, the deep learning technique is more crucial for predicting crop yields from the remote sensing images, and more complicated approaches are needed to derive the essential spatiotemporal features of the data. Thus, this paper suggests a new weight-optimized Stacked Long Short Term Memory with Conditional Random Field (SLSTM-CRF) and a Modified Self-Adaptive Generalized Normal Distribution Optimizer (MSGNDO) for the prediction of crop yield. This research work has the following phases; data collection, data preprocessing, weighted feature selection, and crop yield forecasting. Primarily, the standard agricultural data is taken from the benchmark sources. Then, data preprocessing is carried out to improve the quality of data. Next, the features are optimally chosen using a newly recommended MSGNDO, in which the weights are tuned via MSGNDO. Further, the tuned weights are multiplied by the extracted features. The final crop yield prediction is done via weight-optimized SLSTM-CRF, where the parameter optimization is done using the same MSGNDO method. The prediction performance of the suggested framework is compared with existing techniques using diverse error-based measures to show effective performance regarding Dataset 1 and Dataset 2. While considering the Root Mean Squared Error (RMSE) analysis, the suggested framework shows 22.4571 less than 29.88774 of CNN, 29.20486 of RNN, 27.9877 of ResNet-50, and 25.38448 of LSTM. Overall, the performance analysis of the developed model while using dataset 1 achieves 3.549 and 24.46 for Mean Absolute Error (MAE) and RMSE, respectively. While comparing dataset 2, the developed model shows values of 5.196 and 29.16 regarding MAE and RMSE measures. Hence, the entire experimental evaluation of the recommended methods shows more effective outcomes than the existing conventional models.}, archive = {J_ASOC}, author = {Shaik Shameer Basha and B S Nissar Begum}, doi = {10.1016/j.asoc.2025.113760}, journal = {Applied Soft Computing}, month = {12}, pages = {113760}, shortjournal = {Appl. Soft. Comput.}, title = {Weight optimized stacked LSTM with conditional random fields using self-adaptive generalized normal distribution optimizer for crop yield forecasting}, volume = {184}, year = {2025}, }
• (2025). An optimized logic mining method for data processing through higher-order satisfiability representation in discrete hopfield neural network. ASOC, 184, 113759. (www)

  @article{ , abstract = {A high performance classification tool such as logic mining has emerged as one of the future computing systems for big data processing. The collaboration between logic and neural network resulted in extracting the most suitable induced logic to represent knowledge from real-life datasets. However, there are certain limitations within the current logic mining models including a non-flexible logical structure, non-optimal computation of the best logic, and the generation of overfitting solutions. Motivated by these limitations, a novel logic mining model incorporating the non-systematic Satisfiability, namely Random 3 Satisfiability in Discrete Hopfield Neural Network is proposed as a logical structure to represent the behaviour of the dataset. The proposed logic mining models used flexible logical structures to prevent overfitting solutions and optimize synaptic weight values. A new computational approach of the best logic by considering True Positive and True Negative values of the learning system is applied in this work to preserve the significance behaviour of the dataset. Furthermore, the comparative experiments of the logic mining models by utilizing various repository real-life datasets are conducted from repositories to assess their efficiency. In accordance with the results, the proposed logic mining model dominates in all the metrics for the average rank. The average rank for each metrics are Accuracy (1.9375), Precision (1.9375), Specificity (1.8125), Mathews Correlation (1.5625), and Fowlkes Mallows Index (2.3125). Numerical results and in-depth analysis demonstrate that the proposed logic mining model consistently produces optimal induced logic that best represents the real-life dataset for all the performance metrics used in this study.}, archive = {J_ASOC}, author = {Nurul Atiqah Romli and Nur Fariha Syaqina Zulkepli and Mohd Shareduwan Mohd Kasihmuddin and Syed Anayet Karim and Siti Zulaikha Mohd Jamaludin and Nur ‘Afifah Rusdi and Gaeithry Manoharam and Mohd. Asyraf Mansor and Nur Ezlin Zamri}, doi = {10.1016/j.asoc.2025.113759}, journal = {Applied Soft Computing}, month = {12}, pages = {113759}, shortjournal = {Appl. Soft. Comput.}, title = {An optimized logic mining method for data processing through higher-order satisfiability representation in discrete hopfield neural network}, volume = {184}, year = {2025}, }
• (2025). Enhancing nonlinear dependencies of mamba via negative feedback for time series forecasting. ASOC, 184, 113758. (www)

  @article{ , abstract = {Mamba is a rising model designed to distill complex patterns from historical data, providing predictive capabilities for time series forecasting tasks. Mamba’s similarity to linear-based models has been criticized due to its limited ability to capture nonlinear dependencies. In this work, we propose a novel model named Embedding C hannel Attention M aclaurin E instein Mamba (CME-Mamba 1 .) based on Mamba framework, with both Embedding Channel Attention and Maclaurin mechanisms incorporated. To further address gradient vanishing issues, we integrate Einstein FFT algorithms, ensuring robust performance against abnormal behaviors of Mamba-based architectures. Extensive experiments conducted on 11 real-world datasets with different numbers of variates, domain focus and granularity, reveal that CME-Mamba achieves state-of-the-art performance in both MSE and MAE, while maintaining reasonable memory efficiency and low time cost. The robustness and credibility of all results are substantiated by a comprehensive convergence and stability analysis. Statistically, consolidated by the Friedman Nonparametric Test and the Wilcoxon Signed-Rank Test, CME-Mamba ranks the first place with significance over counterparts. In addition, in terms of time and memory analysis, CME-Mamba is among the top three models for time and memory efficiency. Despite this, our results further demonstrate that the main contributor is the Embedding Channel Attention Block, which greatly enhances nonlinear dependencies over datasets. The Einstein FFT Block effectively suppresses gradient vanishing occurrences and contributes considerably to performance improvements, driving CME-Mamba both stable and promising. Moreover, the Maclaurin Block based on negative feedback is asymptotically stable without additional gradient vanishing issues and pioneered in achieving synergies with other blocks and greatly enhances nonlinear dependencies. With enhanced nonlinear dependencies generated from the synergy effect of all the three blocks, CME-Mamba grows excellent to uncover complex paradigms and predict future states in various domains, especially improving the performance for periodic and high-variate situations, such as traffic flow management ( ≈ + 8 % ), electricity predictions( ≈ + 6 % ).}, archive = {J_ASOC}, author = {Sijie Xiong and Cheng Tang and Yuanyuan Zhang and Haoling Xiong and Youhao Xu and Atsushi Shimada}, doi = {10.1016/j.asoc.2025.113758}, journal = {Applied Soft Computing}, month = {12}, pages = {113758}, shortjournal = {Appl. Soft. Comput.}, title = {Enhancing nonlinear dependencies of mamba via negative feedback for time series forecasting}, volume = {184}, year = {2025}, }
• (2025). Sequential unsupervised–supervised learning for clustering time-dependent patterns using ellipsoidal calculus. ASOC, 184, 113757. (www)

  @article{ , abstract = {This work introduces a sequential clustering algorithm that combines unsupervised and supervised learning methodologies using ellipsoidal calculus and recurrent neural networks (RNNs). The unsupervised clustering algorithm (ULCA) identifies ellipsoidal sets for the data by applying Lagrange multipliers. These sets are then optimized with gradient descent to adjust their volume and orientation. For time-dependent data, the optimized ellipsoidal sets are updated dynamically by an RNN, which refines their center, orientation, and axis sizes in response to changes in the data. The ULCA is compared to density-based spatial clustering (DBSCAN) and K-means algorithms, showing superior accuracy without the need for pre-determined cluster numbers. Additionally, the convergence of the ULCA and RNN algorithms, working sequentially, is formally proven using Lyapunov stability theory, ensuring continuous classification of data that evolves over time. This study demonstrates the advantages of the proposed hybrid method over traditional clustering algorithms.}, archive = {J_ASOC}, author = {Alejandro Guarneros and Mariana Ballesteros and Iván Salgado and Isaac Chairez}, doi = {10.1016/j.asoc.2025.113757}, journal = {Applied Soft Computing}, month = {12}, pages = {113757}, shortjournal = {Appl. Soft. Comput.}, title = {Sequential unsupervised–supervised learning for clustering time-dependent patterns using ellipsoidal calculus}, volume = {184}, year = {2025}, }
• (2025). Hybrid quantum annealing for large-scale exam scheduling: Validation in real-world educational scenarios. ASOC, 184, 113756. (www)

  @article{ , abstract = {This study applied the hybrid quantum annealing to address complex exam scheduling challenges in large-scale educational scenarios through hybrid quantum annealing enhanced by graph-based preprocessing. By formulating the problem as a Quadratic Unconstrained Binary Optimization (QUBO) model and leveraging the D-Wave Advantage system, our method integrates quantum annealing with classical preprocessing to resolve constraints on exam room availability, student-course conflicts, and batch synchronization. The approach was applied in a university’s 2022 pandemic-era makeup exam scheduling for 1807 students and 215 courses (2749 exam instances) with zero conflicts. Experimental results show that hybrid quantum annealing consumes merely 86 ms of Quantum Processing Unit (QPU) execution time. In contrast, classical simulated annealing requires 13547 ms of Central Processing Unit (CPU) execution time for the same problem scale. This work bridges quantum computing and educational operations, offering a comparative analysis of hybrid algorithms in multi-constraint optimization domains.}, archive = {J_ASOC}, author = {Ziyu Zhou and Qing Chen and Chaojie Zhang and Mengtong Tan and Shuyan Li}, doi = {10.1016/j.asoc.2025.113756}, journal = {Applied Soft Computing}, month = {12}, pages = {113756}, shortjournal = {Appl. Soft. Comput.}, title = {Hybrid quantum annealing for large-scale exam scheduling: Validation in real-world educational scenarios}, volume = {184}, year = {2025}, }
• (2025). Multi-phase evolutionary search space shrinking for large-scale multi-objective feature selection. ASOC, 184, 113755. (www)

  @article{ , abstract = {Feature selection is essential in machine learning, especially for high-dimensional datasets, where irrelevant and redundant features can degrade performance and increase computational cost. In such settings, the search space becomes exponentially large and sparsely populated with relevant solutions, making effective exploration highly challenging. Despite progress in evolutionary methods, many existing algorithms struggle to scale or maintain sparsity. There is an urgent need for scalable strategies that intelligently reduce the search space while retaining essential features. Using feature importance to guide search space shrinking offers a powerful, domain-specific approach tailored for feature selection. This paper proposes a novel large-scale multi-objective evolutionary algorithm, LMSSS, that addresses sparse feature selection through a multi-phase search space shrinking strategy. Features are ranked based on correlation with class labels and frequency in an initial low-cost evolutionary process. A voting-based crossover operator prioritizes parent solutions with better classification accuracy, while a guided mutation mechanism reintroduces prematurely excluded features for reevaluation. These components enable efficient exploration of sparse, high-dimensional spaces. Experiments on 15 large-scale datasets demonstrate that LMSSS achieves superior classification accuracy with smaller, more informative feature subsets compared to state-of-the-art methods, highlighting its effectiveness and scalability.}, archive = {J_ASOC}, author = {Azam Asilian Bidgoli and Shahryar Rahnamayan}, doi = {10.1016/j.asoc.2025.113755}, journal = {Applied Soft Computing}, month = {12}, pages = {113755}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-phase evolutionary search space shrinking for large-scale multi-objective feature selection}, volume = {184}, year = {2025}, }
• (2025). Adaptive information mesh for multimodal brain tumor segmentation. ASOC, 184, 113754. (www)

  @article{ , abstract = {Utilizing advanced deep learning techniques for automatically segmenting brain tumors in multiparametric magnetic resonance imaging (mpMRI) is crucial in enhancing diagnostic processes. However, extreme data imbalance between brain tumors and non-tumorous tissues, as well as among different subregions within the tumors, poses significant challenges for precise tumor segmentation. In this study, we revisit the problem of positive and negative sample balance from the perspective of segmentation difficulty, integrating information entropy theory. We propose an Adaptive Information Mesh that resamples based on the conditional entropy between each pixel in the original data and the final segmentation target. This resampling approach standardizes the data to a uniform level of segmentation difficulty while preserving contextual information. Our methodology includes Multi-density Sparse Convolution and Topological Reconstruction Operator to process data effectively after AI mesh sampling. Multi-density Sparse Convolution exploits the advantages of mesh data to overcome the limitations of traditional sparse convolutions, which cannot exchange information over long distances due to a fixed receptive field. The Topological Reconstruction Operator rebuilds the topological relationships between mesh layers, facilitating information exchange. We conducted training on the BraTS 2018, 2019, and 2021 datasets. Comprehensive benchmark testing on the same metrics demonstrates that our performance is comparable to state-of-the-art structured networks in terms of accuracy while reasonably allocating computational resources. Results on the ISLES 2022 dataset for infarct segmentation in ischemic stroke underscore the robustness of our method.}, archive = {J_ASOC}, author = {Qingfan Hou and Yanjun Peng and Zhuofei Wang and Jian Jiang and Nan Lv}, doi = {10.1016/j.asoc.2025.113754}, journal = {Applied Soft Computing}, month = {12}, pages = {113754}, shortjournal = {Appl. Soft. Comput.}, title = {Adaptive information mesh for multimodal brain tumor segmentation}, volume = {184}, year = {2025}, }
• (2025). An attention-based multiple band interaction network for motor imagery EEG decoding. ASOC, 184, 113750. (www)

  @article{ , abstract = {Brain-computer interfaces (BCIs) based on motor imagery electroencephalogram (MI-EEG) have been extensively used in many applications to assist disabled people. Key frequency bands play a crucial role in MI-EEG signal decoding. However, existing researches pay insufficient attention to the interaction among different bands, especially during feature extraction process. To address this issue, in this work, a multiple band interaction network, called MBINet, is carefully constructed, which adopts band-dependent multi-branch setting. Particularly, an attention-based guide block and a multi-scale interaction block are elaborately designed. The former allows to improve the feature extraction process of μ and β bands by the lights of full band characteristics, and the latter enables to promote multi-angle and multi-scale interactive fusion of high-order significant features from multiple bands. The performance testing of MBINet is performed on two publicly available MI datasets, namely the BCI competition IV-2a dataset and the High gamma dataset. The experimental results show that MBINet outperforms the state-of-the-art methods, with an average classification accuracy of 81.66% and 95.78%, respectively. MBINet provides a new perspective for decoding nonlinear time series, especially EEG signals, by emphasizing diverse interactions between key frequency bands.}, archive = {J_ASOC}, author = {Weidong Dang and Kefa Zhang and Haoyu Li and Dongmei Lv and Wei Guo and Zhongke Gao}, doi = {10.1016/j.asoc.2025.113750}, journal = {Applied Soft Computing}, month = {12}, pages = {113750}, shortjournal = {Appl. Soft. Comput.}, title = {An attention-based multiple band interaction network for motor imagery EEG decoding}, volume = {184}, year = {2025}, }
• (2025). The neighborhood rough set based on division-mining-fusion strategy. ASOC, 184, 113749. (www)

  @article{ , abstract = {The neighborhood rough set (NRS) model exhibits significant value in numerous data mining tasks. However, most existing NRS models are only suitable for data mining problems in a single dataset, and their comprehensive performance in processing data is not ideal. To address these two challenges, we develop a novel NRS model by applying the division-mining-fusion (DMF) strategy. Specifically, we first decompose the original dataset into multiple subsets. Then, a submodel is established for each subset, and finally, all submodels are integrated to develop a new NRS model. The experimental results demonstrate that the developed model can not only effectively handle the data mining tasks in scenarios with multiple datasets, but also has outstanding comprehensive performance. The NRS model developed in the paper provides a novel solution for efficiently processing large-scale complex data.}, archive = {J_ASOC}, author = {Conghao Yan and Qingzhao Kong and Wenbin Zhang}, doi = {10.1016/j.asoc.2025.113749}, journal = {Applied Soft Computing}, month = {12}, pages = {113749}, shortjournal = {Appl. Soft. Comput.}, title = {The neighborhood rough set based on division-mining-fusion strategy}, volume = {184}, year = {2025}, }
• (2025). Explainable deep learning framework for lower limb activity classification using reconstructed sEMG signals in healthy and pathological subjects. ASOC, 184, 113748. (www)

  @article{ , abstract = {Recognizing human activities from Surface Electromyography (sEMG) signals is fraught with challenges, including noise susceptibility and signal crosstalk. Addressing these, our study analyses sEMG data from 11 healthy and 11 pathological subjects across three distinct activities: sitting, standing, and walking. We introduce a pioneering preprocessing approach that integrates Bandpass Filtering, Wavelet Denoising, and Ensemble Empirical Mode Decomposition (EEMD) for signal enhancement. A significant aspect of our approach involves the reconstruction of signals by selecting the top 50% of Intrinsic Mode Functions (IMFs) based on entropy, Signal-to-Noise Ratio (SNR), correlation, and energy, effectively capturing the most informative features of the sEMG signals. To counterbalance dataset imbalances and facilitate robust feature extraction, we applied Adaptive Synthetic (ADASYN) sampling and segmented the data into 256 ms windows with a 25% overlap. Our Convolution Neural Network (CNN) achieves remarkable classification accuracies: for sitting, 99.4% in healthy and 99.2% in pathological subjects; for standing, 99.7% and 98.8% respectively; and for walking, 99.4% and 98.6%, respectively. Furthermore, the integration of Explainable AI (XAI) through Permutation Feature Importance (PFI) provides critical insights into the significant impact of muscle signals, particularly highlighting the Rectus Femoris (RF) muscle’s role in sitting with leg extension, BF muscles’s role in standing with flexion. This comprehensive and innovative methodology not only overcomes the inherent challenges of sEMG signal analysis but also enhances the interpretability and reliability of activity recognition, marking a significant advancement for personalized healthcare interventions.}, archive = {J_ASOC}, author = {Pratibha Tokas and Vijay Bhaskar Semwal and Sweta Jain}, doi = {10.1016/j.asoc.2025.113748}, journal = {Applied Soft Computing}, month = {12}, pages = {113748}, shortjournal = {Appl. Soft. Comput.}, title = {Explainable deep learning framework for lower limb activity classification using reconstructed sEMG signals in healthy and pathological subjects}, volume = {184}, year = {2025}, }
• (2025). Mission assignment and path planning for ground forces using permutation-based simplified swarm optimization with Q-learning adaptation. ASOC, 184, 113747. (www)

  @article{ , abstract = {This study tackles task assignment and path planning for counteroffensive operations using a bi-level programming model inspired by spatial crowdsourcing. While existing models often prioritize centralized platform decisions, they tend to overlook the autonomy and real-time judgment of on-site participants, whose local actions critically affect coordination outcomes. To address this, we propose the Defender Deployment and Routing Model (DDRM), which integrates upper-level task assignment with lower-level route planning. To solve DDRM efficiently, we develop a hybrid nested framework integrating an Adaptive Permutation-Based Simplified Swarm Optimization (APSSO) algorithm with Dijkstra’s algorithm. APSSO incorporates a task-aware initialization strategy and a lightweight Q-learning mechanism for operator selection. As the first SSO-based approach with permutation-based operators, APSSO effectively handles discrete task allocation structures. Experimental results on 36 benchmark instances show that APSSO achieves the best performance in 34 cases compared to six tailored evolutionary variants. Statistical analysis confirms its superiority in solution quality, while the added Q-learning module introduces minimal computational overhead. These results highlight APSSO’s efficiency and robustness across varying problem scales.}, archive = {J_ASOC}, author = {Jun-Lin Lin and Chyh-Ming Lai and Song-Pei Wu}, doi = {10.1016/j.asoc.2025.113747}, journal = {Applied Soft Computing}, month = {12}, pages = {113747}, shortjournal = {Appl. Soft. Comput.}, title = {Mission assignment and path planning for ground forces using permutation-based simplified swarm optimization with Q-learning adaptation}, volume = {184}, year = {2025}, }
• (2025). Dynamic event-triggered adaptive fuzzy admittance control of robotic systems with uncertainties. ASOC, 184, 113746. (www)

  @article{ , abstract = {In modern industrial production and daily life, the increasing complexity of application scenarios has led to higher requirements for the compliance and safety of robotic systems. However, during robot control, a large number of redundant signals are often sampled, which significantly increases the communication burden. Therefore, how to substantially reduce the communication burden while ensuring satisfactory performance has become an urgent issue to be addressed. To address this challenge, this paper proposes a dynamic event-triggered adaptive fuzzy admittance control (DETAFAC) strategy for robotic systems with uncertainties, where the more aggressive dynamic event-triggered condition can significantly reduce the communication burden and the admittance model is used to reshape the desired trajectory of the robotic systems. Additionally, a fuzzy logic system (FLS) is utilized to address the uncertainties of the robotic systems, the update law of the FLS and the stability of the control system are examined using the Lyapunov stability theorem, and the dynamic triggering condition is formulated to prevent Zeno behavior. Simulation and experimental validations are performed, and the results demonstrate that the proposed DETAFAC strategy can achieve better performances in comparison to the similar approaches.}, archive = {J_ASOC}, author = {Jinzhu Peng and Xuxin Liu and Shuai Ding and Yaqiang Liu}, doi = {10.1016/j.asoc.2025.113746}, journal = {Applied Soft Computing}, month = {12}, pages = {113746}, shortjournal = {Appl. Soft. Comput.}, title = {Dynamic event-triggered adaptive fuzzy admittance control of robotic systems with uncertainties}, volume = {184}, year = {2025}, }
• (2025). Boxing action recognition using inertial data and deep learning. ASOC, 184, 113745. (www)

  @article{ , abstract = {This study explores the ongoing shift toward automated sports performance analysis by employing Deep Convolutional Neural Networks (DCNNs) for the classification of movement data in combat sports, specifically boxing. Using data from wearable inertial measurement units (IMUs), this study addresses the limitations of traditional, expensive video analysis by providing an accessible, cost-effective alternative that can be seamlessly integrated into athletes' training routines. IMUs offer a practical solution for continuous performance monitoring and feedback, establishing a foundation for automated performance assessment during training sessions. In this study, the classification accuracy and recall of DCNNs are rigorously compared with alternative algorithms, demonstrating a high overall recall of 99 % and accuracy of 91 %. Focusing on pad work training as a key application area, this work advances the automation of boxing action recognition, with implications for both training optimization and real-time scoring automation in competitive boxing. These findings underscore the potential for wearable technology and advanced machine learning methods to transform athletic performance evaluation and bring engineering innovation to sports training.}, archive = {J_ASOC}, author = {J. Brindha and G. Nallavan and Radana Vilimkova Kahankova and Radek Martinek}, doi = {10.1016/j.asoc.2025.113745}, journal = {Applied Soft Computing}, month = {12}, pages = {113745}, shortjournal = {Appl. Soft. Comput.}, title = {Boxing action recognition using inertial data and deep learning}, volume = {184}, year = {2025}, }
• (2025). Adaptive dual-network control for modular robots in human–robot interaction. ASOC, 184, 113744. (www)

  @article{ , abstract = {In order to address safety and compliance challenges in physical human–robot interaction (pHRI), this paper presents an adaptive dual-neural network (NN) control framework for modular robot manipulators (MRMs). The proposed approach enables seamless switching between autonomous trajectory tracking and compliant motion guided by human interaction. An adaptive fuzzy model uncertainty compensation method is employed to estimate the interaction torque that is utilized to detect physical contact, determining the MRM’s operation mode. Additionally, the reference trajectory tracking information is provided by human motion intention identification via NNs. This method avoids the need for external force/torque sensors and increases the intention recognition accuracy by jointly estimating human impedance parameters. The overall control scheme simultaneously incorporates adaptive fuzzy PD control and decentralized impedance control through a mode-switching structure, ensuring compliant and accurate motion in the two operation modes. A Lyapunov-based analysis is then conducted. The obtained results demonstrate that the tracking errors of the closed-loop system are uniformly ultimately bounded (UUB). An experimental validation for various tasks also demonstrates the high robustness and effectiveness of the proposed strategy. Note that all the procedures adopted in this paper are conducted following standard safety guidelines, with full consideration of ethical aspects in pHRI.}, archive = {J_ASOC}, author = {Yuexi Wang and Tianjiao An and Bo Dong and Mingchao Zhu and Yuanchun Li}, doi = {10.1016/j.asoc.2025.113744}, journal = {Applied Soft Computing}, month = {12}, pages = {113744}, shortjournal = {Appl. Soft. Comput.}, title = {Adaptive dual-network control for modular robots in human–robot interaction}, volume = {184}, year = {2025}, }
• (2025). Dynamic-informed machine learning: Data-driven reconstruction of delay differential equations models. ASOC, 184, 113743. (www)

  @article{ , abstract = {To enhance the understanding of the dynamical mechanisms underlying complex systems, this study presents a novel data-driven framework for modeling Delay Differential Equations. The Galerkin-Koornwinder theory is combined with Neural Ordinary Differential Equations for the first time, addressing the significant challenge of reconstructing models that incorporate time-delay effects by approximating them with Ordinary Differential Equations within a rigorous mathematical framework. Leveraging Neural Ordinary Differential Equations to learn these approximate models, the approach achieves high interpretability and generalizability, effectively capturing the dynamics of various delay systems, including those with discrete delays, multi-time delays, and distributed delays. Extensive simulation experiments on models exhibiting bifurcation and chaos, as well as real-world biological data, demonstrate the superior performance of this method in long-term prediction and system dynamics reconstruction. When validated on a set of real-world biological experimental data, the long-term behavior prediction error was reduced from 11% to below 1%. The comparison with similar studies elucidates that the approach can accurately capture complex behaviors and outperforms existing methods in terms of prediction accuracy and generalization capability. All data and code are openly available, facilitating reproducibility and further research.}, archive = {J_ASOC}, author = {Xiyuan Chen and Zhong Liu and Qiubao Wang and Zikun Han}, doi = {10.1016/j.asoc.2025.113743}, journal = {Applied Soft Computing}, month = {12}, pages = {113743}, shortjournal = {Appl. Soft. Comput.}, title = {Dynamic-informed machine learning: Data-driven reconstruction of delay differential equations models}, volume = {184}, year = {2025}, }
• (2025). Fusion of data-driven models with a knowledge-guided loss function for flood forecasting. ASOC, 184, 113742. (www)

  @article{ , abstract = {Heavy rainfall frequently causes flooding disasters in basins of all sizes; however, small and medium-sized basins -typically defined as drainage areas less than 1000 km²- are particularly vulnerable due to their rapid runoff response, often resulting in casualties and property losses. Researchers worldwide have made excellent progress in advancing flood prediction capabilities by developing intelligent data-driven models. However, these models rely only on available data and ignore expert knowledge of floods and fail to account for the varying importance of prediction errors during flood and non-flood events in their training process. Therefore, we propose fusing data-driven models with expert knowledge frameworks by designing a custom loss function for accurate flood forecasting. The proposed model effectively interprets the complexity of flood events in small and medium-sized basins and introduces accurate results. By incorporating expert knowledge constraints, it emphasises the importance of prediction errors during critical flood events. The model was tested using hourly streamflow data from the Heihe and Tunxi basins in China. Results show that our model improves prediction accuracy by up to 80 % in root mean square error (RMSE) and 87 % in mean absolute error (MAE) in Heihe and increases the coefficient of determination (R²) by 66 % and the Kling–Gupta efficiency (KGE) by 58 % in Tunxi compared to baseline models such as INFORMER, TD-CNN-LSTM, STA-LSTM, STA-TCN, LSTM, CNN, TCN, and SVR. These findings demonstrate the model’s superior performance in providing accurate and reliable 6-hour-ahead (T + 6) flood forecasts.}, archive = {J_ASOC}, author = {Haider Malik and Jun Feng and Mohammed Abdallah and Jiru Zhang and Pingping Shao and Zaid Ameen Abduljabbar}, doi = {10.1016/j.asoc.2025.113742}, journal = {Applied Soft Computing}, month = {12}, pages = {113742}, shortjournal = {Appl. Soft. Comput.}, title = {Fusion of data-driven models with a knowledge-guided loss function for flood forecasting}, volume = {184}, year = {2025}, }
• (2025). Self-organizing interval type-2 fuzzy neural network based on eigenvalue decomposition. ASOC, 184, 113741. (www)

  @article{ , abstract = {In this study, an eigenvalue decomposition-based self-organizing interval type-2 fuzzy neural network (ED-SOIT2FNN) is proposed to tackle the identification problem of nonlinear systems. The network model determines both the structure and parameters of the network through online learning, which can realize the structure learning and parameter learning simultaneously. Firstly, in the structure learning process of ED-SOIT2FNN, the error criterion and the completeness criterion of fuzzy rules are used to verify whether the rules grow. Meanwhile, the eigenvalue decomposition method is adopted to find the less active rules for deletion, so that obtain a more compact network structure. Secondly, in terms of ED-SOIT2FNN parameter optimization and the characteristics of network parameters, they are divided into the linear and nonlinear ones. The algorithm of adaptive discount recursive partial least square is employed to optimize the linear parameters, which is conducive to improving the noise resistance of the network model and solving the data saturation problem. And the sliding window adaptive second-order algorithm with a forgetting factor is adopted to optimize the nonlinear parameters. Compared with the algorithm of gradient descent optimization, it can accelerate the convergence and achieve a good adaptability with stability. Finally, the proposed ED-SOIT2FNN was applied to four typical nonlinear examples for identification. The experimental results showed that compared with similar methods in the existing literature, the proposed ED-SOIT2FNN could produce a more compact network structure with higher accuracies of identification and prediction.}, archive = {J_ASOC}, author = {Panchao Wang and Taoyan Zhao and Jiangtao Cao and Ping Li}, doi = {10.1016/j.asoc.2025.113741}, journal = {Applied Soft Computing}, month = {12}, pages = {113741}, shortjournal = {Appl. Soft. Comput.}, title = {Self-organizing interval type-2 fuzzy neural network based on eigenvalue decomposition}, volume = {184}, year = {2025}, }
• (2025). Berth-quay crane-experiment allocation method based on improved genetic algorithm for cargo and scientific research port. ASOC, 184, 113740. (www)

  @article{ , abstract = {Against the background of booming global shipping trade and growing demand for marine scientific research, multifunctional ports with both cargo transportation and scientific research functions, namely CSR_Ports, are facing significant operational scheduling challenges. Due to the intricate mutual influence and resource competition among cargo vessels, research vessels, and marine experiments, the existing method struggles to cope with the diversified demands of port operations. To solve the above challenges, this study considers the differences in infrastructure requirements between cargo vessels and research vessels, designs differentiated work areas, abstracts the berth-quay crane-experiment allocation problem into a two-dimensional packing problem with time window constraints, and establishes a dual-objective berth-quay crane-experiment allocation (BQCEA) model to minimize the vessels' average turnaround time and the average variation of the experiment start time. Then, given the limitations of genetic algorithms, such as insufficient adaptive ability and a tendency to fall into local optimality, an improved algorithm, namely NQRGA, is proposed based on neighborhood search, quantum multi-point crossover, and retention mechanism. Finally, the BQCEA model is combined with the NQRGA to propose a novel solution method for solving the berth-quay crane-experiment allocation problem, referred to as BQCEA_NQRGA. The effectiveness of the BQCEA_NQRGA method is evaluated using real operational data from two CSR_Ports in China. The test results show that, compared with the FCFS method, the BQCEA_NQRGA method achieves at least a 49.79 % performance improvement, and the optimization effect gradually increases to 101 % with the increase in the scheduling scale. Compared with the six advanced algorithms selected in this paper, the proposed NQRGA performs best in all six different instances. Notably, the BQCEA_NQRGA method can generate high-quality allocation schemes for 70 vessels and 14 experiments simultaneously within 100 s.}, archive = {J_ASOC}, author = {Ming-Wei Li and Xiang-Yang Li and Jing Geng and Zhong-Yi Yang and Wei-Chiang Hong}, doi = {10.1016/j.asoc.2025.113740}, journal = {Applied Soft Computing}, month = {12}, pages = {113740}, shortjournal = {Appl. Soft. Comput.}, title = {Berth-quay crane-experiment allocation method based on improved genetic algorithm for cargo and scientific research port}, volume = {184}, year = {2025}, }
• (2025). A novel deep reinforcement learning model based on DDPG considering attention mechanism and combined with GRU network for short-term load forecasting. ASOC, 184, 113739. (www)

  @article{ , abstract = {Short-term load forecasting plays a crucial role in power system operations and energy market efficiency. This paper addresses the challenge of capturing complex nonlinear patterns and temporal dependencies in electricity load data, which traditional forecasting methods often fail to handle effectively. We propose a novel deep reinforcement learning approach that combines attention mechanisms with gated recurrent units within a deep deterministic policy gradient framework (Attention-GRU-DDPG). The key innovation lies in treating load forecasting as a decision-making problem, where the model learns to adapt its predictions based on multivariate inputs including historical load, weather conditions, and electricity prices. Our approach uniquely integrates three complementary components: attention mechanisms for feature prioritization, GRU networks for temporal pattern recognition, and reinforcement learning for adaptive strategy optimization. Extensive experiments on Australian electricity market data demonstrate that our model achieves superior forecasting accuracy compared to ten benchmark methods. The proposed framework offers a practical solution for power system operators requiring accurate 24–168 h ahead load predictions, contributing to more efficient grid management and market operations.}, archive = {J_ASOC}, author = {Xin He and Wenlu Zhao and Zhijun Gao and Licheng Zhang and Qiushi Zhang and Xinyu Li}, doi = {10.1016/j.asoc.2025.113739}, journal = {Applied Soft Computing}, month = {12}, pages = {113739}, shortjournal = {Appl. Soft. Comput.}, title = {A novel deep reinforcement learning model based on DDPG considering attention mechanism and combined with GRU network for short-term load forecasting}, volume = {184}, year = {2025}, }
• (2025). A novel blood supply management model using evolutionary neural network. ASOC, 184, 113738. (www)

  @article{ , abstract = {Increasing demand and supply uncertainties and the short lifetime of blood products cause wastage of the blood gathered from donors. Also, there is a high shortage of blood products because of the limited number of donors and emergency demands. Hence, there is a great significance to constructing a blood supply management model by minimizing (1) transportation cost, (2) shortage cost, (3) wastage cost, (4) ordering cost, and (5) inventory holding cost. The different types of costs with constraints make blood supply management a challenging multi-objective combinatorial optimization problem. Earlier, neural genetic algorithms mostly concentrate on single-objective optimization problems. In this paper, we propose a blood supply management procedure using a network-driven evolutionary network (D2NNEA). A set of pointer networks are utilized to solve the multi-objective problem. The performance of the proposed technique is analyzed by comparing it with existing techniques and the superiority of the proposed technique in solving multi-objective optimization problems is verified.}, archive = {J_ASOC}, author = {Harinandan Tunga and Soumyadip Dhar and Samarjit Kar and Debasis Giri}, doi = {10.1016/j.asoc.2025.113738}, journal = {Applied Soft Computing}, month = {12}, pages = {113738}, shortjournal = {Appl. Soft. Comput.}, title = {A novel blood supply management model using evolutionary neural network}, volume = {184}, year = {2025}, }
• (2025). Partial unified learning for dynamic change detection in hyperspectral images. ASOC, 184, 113737. (www)

  @article{ , abstract = {Hyperspectral image change detection (HSI-CD) aims to identify changes in bi-temporal hyperspectral images (HSIs) captured at different times in the same location. Existing algorithms often overlook the inherent class imbalance in HSI-CD, leading to poor generalization in detecting changes while introducing redundant computation in unchanged regions. This paper introduces a novel mechanism based on Partial Unified Learning for Dynamic Change Detection (PUL-DCD) to address these limitations. Particularly, a novel partial unified learning network is proposed, whose backbone is trained using multiple datasets, whilst the task-specific networks are trained independently with each individual dataset. In so doing, the network can maintain outstanding performance on specific datasets while having strong generalization ability. Furthermore, an innovative dynamic architecture is introduced that distinguishes between easy and hard regions for change detection, thereby optimizing parameter configuration and enhancing detection performance in challenging areas, while mitigating redundancy regarding unchanged information. Experimental results on three datasets show that PUL-DCD is competitive in both accuracy and efficiency.}, archive = {J_ASOC}, author = {Keyun Zhao and Ying Li and Qingping Zheng and Qiang Shen}, doi = {10.1016/j.asoc.2025.113737}, journal = {Applied Soft Computing}, month = {12}, pages = {113737}, shortjournal = {Appl. Soft. Comput.}, title = {Partial unified learning for dynamic change detection in hyperspectral images}, volume = {184}, year = {2025}, }
• (2025). Offline reinforcement learning for job-shop scheduling problems. ASOC, 184, 113736. (www)

  @article{ , abstract = {Recent advances in deep learning have shown significant potential for solving combinatorial optimization problems in real-time. Unlike traditional methods, deep learning can generate high-quality solutions efficiently, which is crucial for applications like routing and scheduling. However, existing approaches like deep reinforcement learning (RL) and behavioral cloning have notable limitations, with deep RL suffering from slow learning and behavioral cloning relying solely on expert actions, which can lead to generalization issues and neglect of the optimization objective. Offline RL addresses these challenges by learning from fixed datasets while leveraging reward signals, making it especially suitable for constrained combinatorial problems where online exploration is impractical. This paper introduces a novel offline RL method designed for combinatorial optimization problems with complex constraints, where the state is represented as a heterogeneous graph and the action space is variable. Our approach encodes actions in edge attributes and balances expected rewards with the imitation of expert solutions. We demonstrate the effectiveness of this method on job-shop scheduling and flexible job-shop scheduling benchmarks, achieving superior performance compared to state-of-the-art techniques.}, archive = {J_ASOC}, author = {Imanol Echeverria and Maialen Murua and Roberto Santana}, doi = {10.1016/j.asoc.2025.113736}, journal = {Applied Soft Computing}, month = {12}, pages = {113736}, shortjournal = {Appl. Soft. Comput.}, title = {Offline reinforcement learning for job-shop scheduling problems}, volume = {184}, year = {2025}, }
• (2025). Surrogate-assisted multi-objective covariance matrix adaptation evolution strategies. ASOC, 184, 113728. (www)

  @article{ , abstract = {Although the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is one of the most preferred evolutionary algorithms for single-objective black-box optimization, its performance is limited in the multi-objective space due to its reliance solely on Gaussian-based mutation without any crossover for offspring generation. To address this limitation, this work proposes a surrogate-assisted multi-objective CMA-ES algorithm with an ensemble of offspring generation schemes. In the proposed algorithm, trial solutions are generated from an ensemble of the standard CMA-ES operator and a Genetic Algorithm-inspired operator. Consequently, the solutions are evaluated on a Gaussian Process-based surrogate model, and the solution with the best Expected Improvement (EI) is selected as the generated offspring. Experiments on the Walking Fish Group (WFG) test suite and 18 benchmark multi-objective Neural Architecture Search (NAS) problems demonstrate that the proposed approach is statistically superior to existing multi-objective CMA-ES variants and other state-of-the-art non-CMA-ES multi-objective algorithms. Specifically, the proposed algorithm achieves a win rate of 79.63% and 77.8% on the WFG and NAS test suites, respectively, against other CMA-ES variants, and demonstrates a 68.8% win rate against state-of-the-art algorithms on the NAS test suite.}, archive = {J_ASOC}, author = {Oladayo S. Ajani and Adeyinka Adedigba and Kalyana C. Veluvolu and Rammohan Mallipeddi}, doi = {10.1016/j.asoc.2025.113728}, journal = {Applied Soft Computing}, month = {12}, pages = {113728}, shortjournal = {Appl. Soft. Comput.}, title = {Surrogate-assisted multi-objective covariance matrix adaptation evolution strategies}, volume = {184}, year = {2025}, }
• (2025). A multi-source domain adaptation approach with learning domain-specific representations for bearing fault diagnosis under limited samples. ASOC, 184, 113727. (www)

  @article{ , abstract = {Bearings, vital to train traction motors, directly influence train safety due to their exposure to harsh environments and alternating loads, leading to various faults. Bearing fault diagnosis is thus critical for ensuring train safety. However, the challenge of data privacy and the scarcity of data make diagnosing faults difficult. Currently, transfer models based on single-source domain adaptation are used to address the above issues. However, data from multiple related domains may exist simultaneously in some scenarios. Therefore, learning fault knowledge from multiple domains and transferring it to the target domain for faults is a difficult problem. In view of this, we propose a multi-source domain adaptation approach with learning domain-specific representations (MDA-LDDSR) for bearing fault diagnosis under limited samples. First, MDA-LDDSR adds a domain-specific feature extractor (MAS-Net) for source and target domain to align the target domain with the distribution. After purifying the features by a partial feature selection strategy, we innovatively construct an intra-class alignment strategy to keep the marginal and conditional distributions of the data in alignment. Then, we utilize the instance-to-domain Mahalanobis distance to explicitly model the similarity between different domains. Finally, this similarity is assigned in the form of weights to multiple domain hypotheses. Superiorly, experiments in real rolling bearing fault diagnosis also show that the model has better convergence while it is characterized by high efficiency and robustness. The proposed approach has a strong advantage to be used for real-time bearing fault diagnosis when noise is added to the environment. Also, it obtains the desired diagnosis results compared to other approaches of the same type.}, archive = {J_ASOC}, author = {Qiang Zhou and Wengang Ma and Yadong Zhang and Jin Guo}, doi = {10.1016/j.asoc.2025.113727}, journal = {Applied Soft Computing}, month = {12}, pages = {113727}, shortjournal = {Appl. Soft. Comput.}, title = {A multi-source domain adaptation approach with learning domain-specific representations for bearing fault diagnosis under limited samples}, volume = {184}, year = {2025}, }
• (2025). Surrounding and context-aware network for individual emotion complement in social networks. ASOC, 184, 113726. (www)

  @article{ , abstract = {Most sentiment analysis focuses on identifying individual emotional shifts related to a specific topic or predicting an individual’s emotional state in the future. However, it overlooks the influence of surrounding factors and contextual emotional information within social networks. Furthermore, it is also important to address the gap in understanding the missing emotions of individuals during critical historical moments. In this paper, we propose a surrounding and context-aware network for individual emotion complement(denoted as SCAEC) to infer missing emotional states at pivotal moments in history. The SCAEC network consists of an encoder layer for emotion embedding, a surrounding-aware layer, a context-aware layer, a feature interactive learning layer and an emotion distribution layer. Additionally, we introduce a decay long short-term memory (LSTM) network (DLSTM) within the SCAEC, which effectively extracts emotional features from individuals during moments of emotional absence and allows for the decay of emotional influence over time. We use DLSTM to extract associated emotional features as global features and apply multi-head attention to capture key emotional features as local features from posts/comments before and after moments of emotional absence. This allows us to obtain surrounding and context features based on the global and local features, respectively. The feature interactive learning layer then combines these features between the surrounding and context features to learn their interaction. Additionally, the emotion distribution layer simulates the emotional interaction between the individual and their surroundings. Finally, we derive the individual’s emotion distribution. Experimental results demonstrate that SCAEC model achieves an F1 score of 67.52% on the X and 68.98% on the Microblog dataset, marking an improvement of 1%–10% compared to baseline methods.}, archive = {J_ASOC}, author = {Sai Kang and YaJun Du and Xianyong Li and Xiaoliang Chen and Chunzhi Xie and Jia Liu and Yan-li Lee}, doi = {10.1016/j.asoc.2025.113726}, journal = {Applied Soft Computing}, month = {12}, pages = {113726}, shortjournal = {Appl. Soft. Comput.}, title = {Surrounding and context-aware network for individual emotion complement in social networks}, volume = {184}, year = {2025}, }
• (2025). Spam bot detection on twitter platform using positional attention based dense convolutional neural network. ASOC, 184, 113725. (www)

  @article{ , abstract = {Nowadays, social media platforms such as Twitter and Facebook play a major role in everyday life because of the incredible opportunities they offer users. Twitter is the most essential social media platform since it allows people to express themselves through tweets. However, because of its popularity among a vast number of users, the Twitter platform is being abused by automated accounts known as bots. Since automated accounts transmit fake news, fake ideas, and fake products, early detection of bots on the Twitter platform is critical. Previously, researchers introduced ineffective methods for identifying social media bots. Positional Attention-based Dense Convolutional Neural Network (PAtt_Dense CNN) is a new deep learning-based spam bot detection framework proposed in this paper. Pre-processing of the incoming data includes stop word removal, tokenization, stemming, n-gram identification, user mention, URL and hashtag removal, vocabulary density, and richness analysis. From the pre-processed images, significant features are extracted utilizing N-gram level vectorizer, TF-IDF vectorizer, character level vectorizer, and Extended Bidirectional Encoder Representations from Transformer (EBERT). After feature extraction, binary water wheel plant optimization is used to select the best characteristics. Finally, spam bot detection and classification are performed using a PAtt_Dense CNN. The performance of the proposed technique is then evaluated by analyzing the performance indicators and comparing them to existing procedures. According to the comparison results, the proposed technique achieved the enhanced outcome in terms of Accuracy, Precision, Recall, and F1-Score of 97.8 %, 97.2 %, 98.3 %, and 99.2 %, respectively.}, archive = {J_ASOC}, author = {Hemal Girishkumar Shah and Hiren Joshi}, doi = {10.1016/j.asoc.2025.113725}, journal = {Applied Soft Computing}, month = {12}, pages = {113725}, shortjournal = {Appl. Soft. Comput.}, title = {Spam bot detection on twitter platform using positional attention based dense convolutional neural network}, volume = {184}, year = {2025}, }
• (2025). Consistency- and ordinal consensus-based multiple scenario model for group decision-making with incomplete probabilistic linguistic preference relations. ASOC, 184, 113722. (www)

  @article{ , abstract = {Incomplete probabilistic linguistic preference relations (InPLPRs), as a practical expression to portray the uncertainty of things, can describe the information on decision makers’ (DMs’) evaluation of things in pairwise comparisons in group decision making with two dimensions simultaneously. This paper investigates a group decision-making (GDM) method with InPLPRs to express the preference information of DMs and establish a consensus reaching mechanism for multiple scenarios. First, to obtain incomplete information, we define the concept of opinion swing neighborhood based on referring to the opinions of others and analyzing the mental behavior of DMs. Furthermore, a missing information estimation model of InPLPRs considering the opinion swing neighborhood is developed to obtain consistent optimal estimates. Secondly, the ordinal consensus index is defined, facilitating a more accurate measure of consensus attainment. Then, we thoroughly explore the classification of decision situations using consistency, consensus, illogical rate, and distinguishing index. Based on this, the corresponding consensus optimization models are proposed for different decision scenarios. Subsequently, the DMs’ weights are presented for aggregating individual opinions with acceptable consistency and consensus into group opinions, and the collective opinions are ranked and selected. Finally, numerical examples, simulation experiments, and comparative analysis demonstrate the proposed method’s applicability, effectiveness, and advantages. The proposed method offers a comprehensive GDM approach based on InPLPRs, which can be applied to various real-world GDM scenarios, demonstrating broad applicability.}, archive = {J_ASOC}, author = {Ran Dang and Peide Liu and Peng Wang and Luis Martínez}, doi = {10.1016/j.asoc.2025.113722}, journal = {Applied Soft Computing}, month = {12}, pages = {113722}, shortjournal = {Appl. Soft. Comput.}, title = {Consistency- and ordinal consensus-based multiple scenario model for group decision-making with incomplete probabilistic linguistic preference relations}, volume = {184}, year = {2025}, }
• (2025). Equidistant deep embedding-based multi-label group activity recognition with dependency-constrained training. ASOC, 184, 113721. (www)

  @article{ , abstract = {Group activities in real-world scenarios are often complex and require multiple labels for adequate representation, as they involve multifaceted social interactions among individuals within a small group setting. Existing group activity recognition (GAR) methods primarily focus on model inference, overlooking the inherent challenges of multi-label learning (MLL). One common approach to MLL is the label powerset (LP) method, which transforms multi-hot label vectors into one-hot vectors, converting the multi-label problem into a single-label multi-classification problem. However, this projection loses correlations among active elements in multi-hot vectors, hindering the capture of inter-activity dependencies. To address this limitation, we propose a novel equidistant deep embedding-based (EDE) multi-label GAR framework with dependency-constrained training, building upon the LP approach. Our framework leverages a deep embedding network with self-supervised equidistant regularization loss to project multi-hot label vectors into evenly spaced dense vectors, which not only facilitate problem transformation but also contain latent activity patterns, including label correlations. We treat these dense vectors as new labels and design a corresponding deep learning and classification (DLC) strategy that optimizes all GAR model parameters except for the classification layers. Additionally, we propose a training-only auxiliary branch, dubbed TransOvR module, to bolster the model’s capacity for inter-activity dependency reasoning. By leveraging Transformer’s context-aware capability, this module facilitates interactions between multiple binary classifiers, focusing on correlations between activities. Our extensive experiments demonstrate the effectiveness of our method, outperforming state-of-the-art multi-label GAR methods.}, archive = {J_ASOC}, author = {Lindong Li and Linbo Qing and Pingyu Wang and Yang Xiao and Wang Tang and Yonghong Peng}, doi = {10.1016/j.asoc.2025.113721}, journal = {Applied Soft Computing}, month = {12}, pages = {113721}, shortjournal = {Appl. Soft. Comput.}, title = {Equidistant deep embedding-based multi-label group activity recognition with dependency-constrained training}, volume = {184}, year = {2025}, }
• (2025). Advanced nonlinear optimization of low- and high-resolution medical images using adaptive deep spearman correlation analysis (D-SCA) for pattern sequence recognition. ASOC, 184, 113720. (www)

  @article{ , abstract = {This paper presents the state-of-art model for those healthcare multimedia datasets that have nonlinear limitations during the processing of multi-dimensional or non-linear data that usually decrease the accuracy rate and increase computational cost as per computational resources. Such as correlation among the low-resolution images to high-resolution images, and videos, the reason is non-linearity between data, non-linear dimensions or variation and illusion in the data. The proposed state-of-the-art model uses multiple techniques with transfer learnings approach to solve these issues. This paper further explains the details of the proposed deep convolutional neural network model for visual feature maps extraction and then followed by the proposed algebraic extension of Spearman correlation analysis to create the relationship of low-resolution image to high-resolution image with adaptive optimization. Then it leads to the next component of model as Radial Basis Function Network (RBFN) for non-linear mapping among LR to HR images, for classification the hybrid Xception deep learning model is used and implemented on four benchmark datasets.}, archive = {J_ASOC}, author = {Muhammad Saddam Khokhar and Misbah Ayoub and Zakria and Abdullah Lakhan}, doi = {10.1016/j.asoc.2025.113720}, journal = {Applied Soft Computing}, month = {12}, pages = {113720}, shortjournal = {Appl. Soft. Comput.}, title = {Advanced nonlinear optimization of low- and high-resolution medical images using adaptive deep spearman correlation analysis (D-SCA) for pattern sequence recognition}, volume = {184}, year = {2025}, }
• (2025). Cross-modal deep interaction and modal-aware aggregation network for visible and infrared tracking. ASOC, 184, 113719. (www)

  @article{ , abstract = {Currently, developing a robust RGB-Thermal (RGBT) tracking method in complex environments remains challenging in effectively mining and fusing cross-modal complementary information, enhancing modal perception capabilities, and improving the representation ability of fused semantic features. To address these challenges, we propose a novel architecture called Multi-branch Cross-modal Deep Interaction Fusion and Adaptive Aggregation Integrating Hybrid Attention Semantic Enhancement Network (MCFTNet). Specifically, MCFTNet first extracts modality-specific features through dedicated branches, followed by designing a cross-modal deep interaction fusion network to achieves deep interaction and comprehensive fusion of cross-modal features through a fusion branch as the medium. Furthermore, a modal-aware adaptive aggregation module is developed to dynamically aggregates high-resolution features from different branches, while significantly enhancing the discriminative ability of multi-modal features. Finally, a hybrid attention semantic enhancement module is introduced that combines carefully designed enhanced multi-head attention and hierarchical attention to optimize the fused semantic features, thereby achieving highly accurate prediction of target position and shape. Extensive experiments on three mainstream public benchmark datasets, GTOT, RGBT234, and LasHeR, demonstrate the effectiveness and robustness of our proposed method.}, archive = {J_ASOC}, author = {Xiang Liu and Haiyan Li and Victor Sheng and Yujun Ma and Xiaoguo Liang and Guanbo Wang}, doi = {10.1016/j.asoc.2025.113719}, journal = {Applied Soft Computing}, month = {12}, pages = {113719}, shortjournal = {Appl. Soft. Comput.}, title = {Cross-modal deep interaction and modal-aware aggregation network for visible and infrared tracking}, volume = {184}, year = {2025}, }
• (2025). Embedding neural sampling and adversarial bandit into gene expression programming for symbolic regression. ASOC, 184, 113718. (www)

  @article{ , abstract = {Evolutionary computation methods for symbolic regression problems use a search method to actively explore the mathematical expression space to find a solution. However, the evolutionary computation search process lacks a clear direction due to the randomness of evolutionary operations. In contrast, deep learning methods focus on learning a clear mapping from a given dataset to a mathematical function, but the learning process does not generally involve an active search of the potential solution space. To combine the advantages of the active search of evolutionary computation and the clear mapping direction of neural networks, this paper proposes a novel algorithm called GVAE-ABGEP. GVAE-ABGEP incorporates a grammar variational autoencoder and adversarial bandit into gene expression programming to guide its search process. GVAE-ABGEP partitions the mathematical expression space into many subspaces. It then leverages an adversarial bandit — AvgHExp3 to choose a subspace. In the selected subspace, it then utilizes an autoencoder to sample individuals whose fitnesses are near the local optima, and executes crossover, mutation, and selection in evolutionary computation to explore the global optima. Experiments on 18 symbolic regression and 12 physics benchmarks show that GVAE-ABGEP outperforms three baseline gene expression programming methods and six baseline machine learning methods.}, archive = {J_ASOC}, author = {Qiang Lu and Qiuchen Yuan and Dawei Li and Congwen Xu and Jake Luo and Zhiguang Wang}, doi = {10.1016/j.asoc.2025.113718}, journal = {Applied Soft Computing}, month = {12}, pages = {113718}, shortjournal = {Appl. Soft. Comput.}, title = {Embedding neural sampling and adversarial bandit into gene expression programming for symbolic regression}, volume = {184}, year = {2025}, }
• (2025). Graph-based analysis of problem space and genetic programming classifier performance using optimal transport dataset distance. ASOC, 184, 113716. (www)

  @article{ , abstract = {Understanding and interpreting machine learning tasks is a non-trivial endeavor. To this end, many works have proposed creative ways to organize and visualize a problem dataset, the learning process itself, or the models produced after learning has occurred. Moreover, as the scope of machine learning increases, the re-purposing or specialization of models becomes more common, or as continued automation is sought for algorithm deployment, understanding sets of problems has also become relevant. The present work proposes a graph-based representation of a set of machine learning problems, a novel characterization of the relationships and structures that are present in such a set. While similar proposals have been made before, the proposed methodology employs a recent and unique metric between problem datasets, the Optimal Transport Dataset Distance, which allows for the computation of a mathematically rigorous distance matrix for problem sets. This allows for the construction of a graph-based representation, while previous works relied on representations based on problem meta-features that were defined heuristically. Results show that the resulting graph representation of a problem set exhibits structural properties that are related to empirical indicators of problem difficulty, such as the average error, the size of the dataset, and the class imbalance. A similar analysis using meta-features shows that the structure of the resulting graphs cannot capture the same nuanced relationships between problems.}, archive = {J_ASOC}, author = {Joel Lee Nation and Daniel Fajardo and Yuliana Martínez and Arnoldo Díaz-Ramírez and Leonardo Trujillo}, doi = {10.1016/j.asoc.2025.113716}, journal = {Applied Soft Computing}, month = {12}, pages = {113716}, shortjournal = {Appl. Soft. Comput.}, title = {Graph-based analysis of problem space and genetic programming classifier performance using optimal transport dataset distance}, volume = {184}, year = {2025}, }
• (2025). Graph-based deep reinforcement learning for dynamic scheduling of flexible job-shop considering worker fatigue and multi-skill factors. ASOC, 184, 113712. (www)

  @article{ , abstract = {In recent years, the dynamic scheduling problem of flexible job shops (DFJSP) has garnered significant attention. However, most DFJSP studies focus solely on machine constraints, often overlooking the crucial factor of worker constraints. As an essential resource within the production process, the efficient utilization of workers can substantially enhance production efficiency. Consequently, the scheduling problem in dual-resource constrained (DRC) production systems, where operation times fluctuate according to worker skill levels and fatigue, is investigated. Considering new job arrivals as dynamic events, this study comprehensively integrates worker skill and fatigue factors, with the goal of minimizing total cost and makespan by developing a dual-resource constrained dynamic flexible job-shop scheduling (DRC-DFJSP) optimization model. To enable real-time model resolution in dynamic environments, an end-to-end deep reinforcement learning (DRL) scheduling method is introduced. The decision-making process is formulated as a Markov decision process (MDP) and guided by a reward mechanism tailored to optimization objectives. An enhanced proximal policy optimization (PPO) algorithm, combined with an adaptive clipping mechanism and a prioritized experience replay buffer, is applied to handle operation ordering, machine assignment, and worker allocation within the DRC-DFJSP framework. To further enhance decision-making capabilities, an attention-based graph neural network feature extraction method is incorporated to capture the intricate connections between operations, machines, and workers, resulting in a more precise characterization of the workshop state. Numerical experiments and case studies demonstrate that the proposed scheduling method surpasses existing strategies in dynamic environments that account for worker skill and fatigue levels.}, archive = {J_ASOC}, author = {Yiwen Hu and Zequn Zhang and Jie Chen and Dunbing Tang and Qixiang Cai}, doi = {10.1016/j.asoc.2025.113712}, journal = {Applied Soft Computing}, month = {12}, pages = {113712}, shortjournal = {Appl. Soft. Comput.}, title = {Graph-based deep reinforcement learning for dynamic scheduling of flexible job-shop considering worker fatigue and multi-skill factors}, volume = {184}, year = {2025}, }
• (2025). Command and control network architecture optimization method based on dual-layer weighting and TOPSIS-GRA of indicators. ASOC, 184, 113711. (www)

  @article{ , abstract = {In modern warfare, the effectiveness of command and control (C2) networks plays a pivotal role in shaping operational performance and determining combat outcomes. Even under identical resource allocations, variations in command strategies and network architectures can lead to vastly different results. While many existing optimization methods rely on complex network theory to enhance system robustness, they often exhibit randomness and overlook commanders’ intent and expert knowledge during architecture design. Unlike civilian networks, military C2 networks are highly dependent on operational planning and domain expertise, emphasizing decision transparency and interpretability. Subjective evaluation methods-such as those based on rules or expert logic-are more intuitive for commanders but suffer from bias and inconsistency. Conversely, purely data-driven models like deep learning lack transparency, limiting their practical utility in high-stakes scenarios. To bridge this gap, this study proposes a dual-layer weighting optimization approach based on the TOPSIS and Grey Relational Analysis (GRA) methods. First, C2 network architectures are constructed using complex network theory. Then, an evaluation index system is developed around four key metrics: invulnerability, communication efficiency, connectivity, and network density. Subjective weights are derived via the Analytic Hierarchy Process (AHP), while objective weights are calculated using an improved entropy method. These are integrated into a variable-weight model to obtain final weights for optimization. By combining TOPSIS and GRA, the method accounts for both subjective preferences and objective measurements, ensuring a balanced evaluation. Simulation results demonstrate that the proposed method comprehensively considers both subjective and objective factors as well as the impact of state changes on weight adjustments, aligning more closely with real battlefield environments. The constructed C2 network architecture can adapt to different combat missions, guided by subjective evaluations of commander intent and expert experience while being refined through objective assessments to achieve adaptive adjustments. The ranking results account for both the distance and shape variations among indicators. Statistical analysis and comparisons with other methods further validate the effectiveness of the proposed approach, which also reduces computation time by a factor of five compared to alternative algorithms, highlighting its superiority.}, archive = {J_ASOC}, author = {Jianwei Wang and Qing Zhang and Chengsheng Pan}, doi = {10.1016/j.asoc.2025.113711}, journal = {Applied Soft Computing}, month = {12}, pages = {113711}, shortjournal = {Appl. Soft. Comput.}, title = {Command and control network architecture optimization method based on dual-layer weighting and TOPSIS-GRA of indicators}, volume = {184}, year = {2025}, }
• (2025). A distributed learning framework with blockchain and privacy-preserving for IoV. ASOC, 184, 113710. (www)

  @article{ , abstract = {The Internet of Vehicles (IoV), as a critical component of the Internet of Things (IoT), constructs a distributed system comprising vehicles, roadside units (RSUs), and cloud servers. In the IoV environment, the secure sharing of data and the protection of privacy are of paramount importance, as they directly impact the decision-making processes of intelligent vehicles and overall road safety. Given the openness of IoV, it faces risks of privacy leakage and poisoning attacks during data exchange and model training, which threaten the integrity and reliability of the data. To address these challenges, this study proposes a privacy protection framework that integrates blockchain technology and differential privacy. This framework incorporates dual differential privacy techniques within federated learning to enhance data privacy protection and designs a dynamic gradient aggregation mechanism to defend against data poisoning attacks, thereby ensuring the security of the data. Experimental results demonstrate that this framework maintains high model accuracy even under attack rates of up to 30%, exhibiting remarkable resilience against such attacks. Overall, this study emphasizes the significance of data security and privacy protection in the IoV domain and illustrates the potential of blockchain and differential privacy technologies in enhancing the security of IoV data and safeguarding user privacy. This research provides robust support for the sustainable development of IoV.}, archive = {J_ASOC}, author = {Chunhai Li and Yan Long and Yong Ding and Changsong Yang and Chuan Zhang and Meng Shen and Liehuang Zhu}, doi = {10.1016/j.asoc.2025.113710}, journal = {Applied Soft Computing}, month = {12}, pages = {113710}, shortjournal = {Appl. Soft. Comput.}, title = {A distributed learning framework with blockchain and privacy-preserving for IoV}, volume = {184}, year = {2025}, }
• (2025). Chaotic vortex nanoparticle swarms (CVNPS): Formation and stability analysis. ASOC, 184, 113708. (www)

  @article{ , abstract = {Vortex Nanoparticle Swarms (VNPS) are a class of self-organizing systems that exhibit vortex-like collective behavior but often suffer from dynamic instability and poor angular momentum conservation. While VNPS may achieve partial spatial stability, coherent motion is typically unsustained. To address this, the Chaotic Vortex Nanoparticle Swarm (CVNPS) model is proposed, incorporating chaotic perturbations via logistic maps within a discrete particle model governed by the Generalized Morse Potential. This integration enhances system stability by accelerating convergence, reducing energy fluctuations, and promoting sustained vortex formation. Comparative analysis using key performance indicators – average nearest neighbor distance, polarization, angular momentum, swarm diameter, total energy, energy dissipation rate, and variance – demonstrates that CVNPS significantly outperforms traditional VNPS. It achieves faster stabilization and maintains coherent structure over time, even under varied initial conditions. By addressing both configurational and dynamical stability, CVNPS establishes a foundation for more reliable nanoparticle swarm applications in nanotechnology.}, archive = {J_ASOC}, author = {Mahvish Khurshid Bijli and Prabal Verma and Amrit Pal Singh}, doi = {10.1016/j.asoc.2025.113708}, journal = {Applied Soft Computing}, month = {12}, pages = {113708}, shortjournal = {Appl. Soft. Comput.}, title = {Chaotic vortex nanoparticle swarms (CVNPS): Formation and stability analysis}, volume = {184}, year = {2025}, }
• (2025). Aspect-based context iterative network with hyperbolic tangent function for aspect-based sentiment classification. ASOC, 184, 113707. (www)

  @article{ , abstract = {In the domain of natural language processing, sentiment analysis emerges as a pivotal undertaking. Especially, aspect-based sentiment classification has garnered significant attention in recent years, as it can identify and evaluate emotions related to specific aspects within sentences. Existing approaches typically achieve satisfactory results by extracting keyword information from the context to identify polarity. However, a common challenge faced by these methods is the inclusion of irrelevant words in the extracted keywords, leading to decreased classification accuracy. To tackle this issue, we propose an aspect-based context multiple iteration approach, which leverages the correlation between aspects and context and employs the multi-head attention mechanism to iteratively extract contextual keywords. By doing so, we aim to mitigate the interference of irrelevant words and enhance the accuracy of sentiment classification. Additionally, we address the peculiarities of hard samples by introducing a novel loss function that cleverly incorporates the hyperbolic tangent function and allows for improved model accuracy. To validate the effectiveness of our proposed approach, we conduct extensive experiments on four widely datasets and demonstrate the efficacy of our model in improving sentiment classification.}, archive = {J_ASOC}, author = {Chao Zhu and Benshun Yi and Laigan Luo}, doi = {10.1016/j.asoc.2025.113707}, journal = {Applied Soft Computing}, month = {12}, pages = {113707}, shortjournal = {Appl. Soft. Comput.}, title = {Aspect-based context iterative network with hyperbolic tangent function for aspect-based sentiment classification}, volume = {184}, year = {2025}, }
• (2025). MOGA: Multi-objective genetic algorithm to select stacking ensemble learning for classification. ASOC, 184, 113706. (www)

  @article{ , abstract = {Stacking Ensemble Learning (SEL) has been effectively integrated with Multi-Objective Optimisation (MOO) heuristics for classification tasks across various domains, including finance, healthcare, and cybersecurity. This study aims to address the challenge of generalising SEL to a diverse set of classification cases. Thus, the Multi-Objective Genetic Algorithm (MOGA) framework is proposed, utilising a Genetic Algorithm (GA) to evolve a population of distinct SELs, each built from a varied subset of base models. The goal is to select the subset that composes the most effective SEL for a given classification task. MOGA is designed with two main objectives—maximising p r e c i s i o n and r e c a l l —which helps to maintain independence from any specific classification case. In addition, incorporating models of varied types ensures adaptability and high performance in different situations. Comprehensive experimentation was conducted on 23 diverse datasets, where MOGA demonstrated high performance in nearly all datasets, outperforming other ensemble learning (EL) methods in 100% of the datasets in p r e c i s i o n , 78% in r e c a l l , 69.5% in f 1 − s c o r e , and 78% in a c c u r a c y . A t-test analysis yielded results of p -value < 0 . 05 , indicating a statistically significant improvement in the a c c u r a c y of the MOGA over the base models. Moreover, the framework’s application can be extended to regression tasks as well.}, archive = {J_ASOC}, author = {Abdellah Rezoug and Mohamed Bader-el-den}, doi = {10.1016/j.asoc.2025.113706}, journal = {Applied Soft Computing}, month = {12}, pages = {113706}, shortjournal = {Appl. Soft. Comput.}, title = {MOGA: Multi-objective genetic algorithm to select stacking ensemble learning for classification}, volume = {184}, year = {2025}, }
• (2025). Deep reinforcement learning with evolutionary algorithm-guided imitation for capacitated vehicle routing problems. ASOC, 184, 113705. (www)

  @article{ , abstract = {The capacitated vehicle routing problem (CVRP) is a complex combinatorial optimization challenge that seeks to determine cost-effective routes for customer deliveries while adhering to specific capacity constraints. Although deep reinforcement learning (DRL) has shown promise in addressing CVRP, it often encounters issues such as slow convergence and suboptimal accuracy. This study introduces an innovative approach that enhances both convergence efficiency and solution quality by integrating DRL with imitation learning (IL), utilizing an evolutionary algorithm (EA) as an expert. The proposed methodology incorporates an attention mechanism-based neural network to effectively capture the intricate features of CVRP. It leverages IL to use EA-generated solutions as expert demonstrations, thereby guiding the DRL model toward a more efficient exploration of the solution space. The REINFORCE algorithm with baseline is employed to ensure stable and rapid training of the DRL model. Experimental results indicate that this hybrid approach significantly outperforms widely adopted baseline methods and approaches the performance levels of advanced algorithms like LKH3. Furthermore, the method demonstrates robust generalization capabilities across various CVRP instances, underscoring its potential for practical applications in diverse routing scenarios. This research contributes to the field by demonstrating how integrating EA as experts within an IL framework can effectively enhance DRL for solving CVRP.}, archive = {J_ASOC}, author = {Wenqiang Zhang and Xiaomeng Wang and Yashuan Mu and Miaolei Deng and Peng Li}, doi = {10.1016/j.asoc.2025.113705}, journal = {Applied Soft Computing}, month = {12}, pages = {113705}, shortjournal = {Appl. Soft. Comput.}, title = {Deep reinforcement learning with evolutionary algorithm-guided imitation for capacitated vehicle routing problems}, volume = {184}, year = {2025}, }
• (2025). A focal quotient gradient system method for deep neural network training. ASOC, 184, 113704. (www)

  @article{ , abstract = {This paper introduces a novel approach for training deep neural networks, leveraging a mini-batch focal quotient gradient system (QGS-Focal). The proposed method addresses critical challenges in imbalanced datasets and optimization efficiency. By introducing residual constraints into the loss function, we construct a quotient gradient system that effectively mitigates model overfitting and gradient vanishing problems. By incorporating a focal loss mechanism, it innovatively solves data imbalance issues at the algorithmic level. The adoption of a mini-batch strategy and limited memory method significantly reduces computational costs. Our comprehensive experiments on imbalanced CIFAR-10 and CIFAR-100 have demonstrated the superiority of QGS-Focal, achieving 83.4 % precision, 83.6 % recall, and 82.9 % F1-score on CIFAR-10, significantly outperforming SGD, Adam, and QGS. Moreover, our approach reduces training time by 13.3 %, enhancing computational efficiency while maintaining superior classification performance. The t-SNE visualization further confirms that QGS-Focal has superior convergence properties compared to traditional optimization approaches.}, archive = {J_ASOC}, author = {Caili Lv and Xian-long Lv and Zhiyuan Wang and Tianqi Zhao and Wei Tian and Qingqing Zhou and Lin Zeng and Min Wan and Chenghu Liu}, doi = {10.1016/j.asoc.2025.113704}, journal = {Applied Soft Computing}, month = {12}, pages = {113704}, shortjournal = {Appl. Soft. Comput.}, title = {A focal quotient gradient system method for deep neural network training}, volume = {184}, year = {2025}, }
• (2025). Partial differential equations and machine learning integration for transit-oriented development. ASOC, 184, 113703. (www)

  @article{ , abstract = {The accurate classification of rail transit stations is critical for advancing Transit-Oriented Development (TOD) and promoting sustainable urban growth. This research presents a novel hybrid framework that integrates Partial Differential Equations (PDEs) with Machine Learning (ML) techniques for the classification of rail transit stations. Unlike conventional TOD models, this study applies the heat equation to the Node, Place, Ridership-Time (NPRT) framework, offering a mathematically grounded approach to capture spatial-temporal dynamics in transit systems. This integration represents the first known application of PDE-based physical modeling combined with supervised learning for classifying transit stations within a TOD framework. This approach significantly enhances the model’s interpretability while maintaining competitive prediction accuracy. Through extensive case studies and empirical validation, the PDE-NPRT model demonstrates strong performance, with Mean Squared Error (MSE) values ranging from 0.0075 to 0.0222. Although slightly outperformed by enhanced ML models—such as K-Nearest Neighbors (KNN), Deep KNN (DKNN), and Deep Distributed Neural Networks (DDNN)—which achieve MSEs as low as 0.0034, the PDE-NPRT model offers a more interpretable and theoretically robust alternative. Additionally, the study introduces a multi-layer modeling strategy that combines regression analysis, clustering algorithms, PDEs, and neural networks, further enriching the understanding of ridership patterns and congestion mechanisms. Clustering outcomes are validated through external indices, confirming the alignment of model predictions with real-world site characteristics. This work represents a significant advancement in TOD modeling, offering a robust and explainable tool for urban planners and decision-makers. By bridging advanced mathematical modeling with machine learning, it paves the way for more intelligent, data-driven, and sustainable urban mobility strategies.}, archive = {J_ASOC}, author = {Ahad Amini Pishro and Shiquan Zhang and Alain L’Hostis and Qixiao Hu and Yuetong Liu and Zhengrui Zhang and Van Duc Nguyen and Yongguo Fu and Tianzeng Li}, doi = {10.1016/j.asoc.2025.113703}, journal = {Applied Soft Computing}, month = {12}, pages = {113703}, shortjournal = {Appl. Soft. Comput.}, title = {Partial differential equations and machine learning integration for transit-oriented development}, volume = {184}, year = {2025}, }
• (2025). A dynamic SLAM algorithm based on improved YOLOv9S. ASOC, 184, 113700. (www)

  @article{ , abstract = {Due to the limitations of conventional visual SLAM algorithms and the time-intensive process of semantic segmentation in dynamic environments, this paper proposes a SLAM algorithm that incorporates an enhanced version of YOLOv9S specifically designed for dynamic scenes. The ECASPPELAN and DRepNCSPELAN4 modules have been developed to improve the model’s detection capabilities in dynamic scenes. Secondly, the improved YOLOv9S is integrated into the ORB-SLAM2 framework, which employs bipolar geometry to reject dynamic targets and utilizes a depth separation algorithm to prevent the erroneous rejection of static features. Subsequently, the detection performance of the model was evaluated on the PASCAL VOC dataset, resulting in a 2% improvement in mAP_0.5 and a 1.3% improvement in mAP_0.5:0.95. The tracking performance of the system is examined in the TUM dynamic dataset, where the absolute trajectory root mean square error (RMSE) is reduced by 97.5%, and the relative rotational RMSE is reduced by 92.9% in comparison to ORB-SLAM2 in the high-dynamics scenario. Additionally, the average running rate surpasses 30 Hz while constructing high-quality dense maps in real indoor dynamic environments. This allows the robot to receive more detailed texture information. The experimental results demonstrate that the enhanced algorithm presented in this paper exhibits enhanced accuracy and resilience in dynamic scenarios.}, archive = {J_ASOC}, author = {Qiguang Zhu and Yuchao Zhao and Haofeng Zhang and Weidong Chen}, doi = {10.1016/j.asoc.2025.113700}, journal = {Applied Soft Computing}, month = {12}, pages = {113700}, shortjournal = {Appl. Soft. Comput.}, title = {A dynamic SLAM algorithm based on improved YOLOv9S}, volume = {184}, year = {2025}, }
• (2025). Neural semantic evaluation of blockchain policy tools in china. ASOC, 184, 113699. (www)

  @article{ , abstract = {To explore the differences in policies guiding the development of the blockchain industry in different regions of mainland China, in this study, we developed a compound neural network model comprising bidirectional long short term memory and deep biaffine attention models that analyses the semantic texts of blockchain policies issued by 31 provinces in mainland China. Machine learning models — specifically, term frequency–inverse document frequency and K-means models are used to implement feature selection of the policy text matrix classification after semantic analysis. Finally, this study proposes an innovative policy tools matching approach. We construct a word-topological map for each text category based on semantic relationships. To validate the effectiveness of these tools, we conduct an empirical analysis using a multivariate linear regression model. The results demonstrate that blockchain policy tools significantly promote blockchain innovation in Mainland China.}, archive = {J_ASOC}, author = {Yuxi Zhang and Haifeng Guo and Ke Peng and Hongzhi Wang}, doi = {10.1016/j.asoc.2025.113699}, journal = {Applied Soft Computing}, month = {12}, pages = {113699}, shortjournal = {Appl. Soft. Comput.}, title = {Neural semantic evaluation of blockchain policy tools in china}, volume = {184}, year = {2025}, }
• (2025). Explainable nature-inspired optimization via virtual and actual multi-objective strategies to establish a smart earthquake early warning system. ASOC, 184, 113698. (www)

  @article{ , abstract = {Geosynthetic-reinforced soil (GRS) structures are considered for reducing displacement and providing economical reinforcement solutions. The risk assessment of these structures against earthquakes, based on the prediction of seismic sliding displacement, is a major challenge in this field. Multi-objective optimization is a powerful machine learning tool for selecting efficient features for high-performance forecasting. This research investigates two strategies based on swarm intelligence and genetic programming for a comprehensive evaluation. These frameworks integrate multiobjective optimization algorithms and Newmark methods for utilizing effective physics-informed features. The first strategy is virtual multi-objective (VMO) optimization by applying particle swarm optimization (PSO) based on minimizing one function via variations of other functions. In this approach, the error function, as a computational error object, is minimized versus the nomination of interpretable feature space as a computational cost object through the virtual Pareto front. The second strategy is actual multi-objective (AMO) optimization by exploiting nondominated sorting genetic algorithm II (NSGA-II) based on minimizing several functions simultaneously with two various approaches, including bi-objective and many-objective algorithms through actual Pareto-optimal solutions. In this approach, the computational error value, computational cost value, and computational time value are minimized at the same time. The main novelty of the first technique is low computational complexity, resulting in high speed due to definite search space dimension-based exploration and exploitation to forecast seismic sliding displacement, whereas the major achievement of the second technique is high computational accuracy due to multiobjective structure-assisted exploitation and exploitation. Through numerical validation by employing the Newmark methods, the resultant model predicts the seismic sliding displacement of these structures using two algorithms efficiently. Nevertheless, both strategies have good performance for intelligent forecasting. The actual many-objective optimization algorithm is a more effective switchable machine learning tool based on the proposed adaptable performance index for developing a smart earthquake early warning software that can precisely detect imminent natural hazards.}, archive = {J_ASOC}, author = {Milad Zarchi and Reza A. Nazari and Kong Fah Tee}, doi = {10.1016/j.asoc.2025.113698}, journal = {Applied Soft Computing}, month = {12}, pages = {113698}, shortjournal = {Appl. Soft. Comput.}, title = {Explainable nature-inspired optimization via virtual and actual multi-objective strategies to establish a smart earthquake early warning system}, volume = {184}, year = {2025}, }
• (2025). DTKD-DL: Dual-teacher knowledge distillation with dual-loops for continuous few-shot relation extraction. ASOC, 184, 113696. (www)

  @article{ , abstract = {This paper introduces a new model named DTKD-DL, designed to address the issue of Continuous Few-Shot Relation Extraction (CFRE) across tasks, with the goal of learning and adapting to newly emerging relations while reducing catastrophic forgetting. In this paper, we have designed a dual-teacher knowledge distillation model based on relation information to enrich knowledge representation and retain prior knowledge. We employ a dual-loops distillation approach, which facilitates knowledge transfer and optimizes the direction of parameter updates, thereby reducing the occurrence of catastrophic forgetting. Furthermore, to avoid overfitting issues caused by multiple rounds of distillation, we have innovatively integrated reinforcement learning with the model. We have validated our model on the FewRel and TACRED datasets and compared it with the large language model Llama3-8b, demonstrating the effectiveness of our model in this scenario and its advantages over the most advanced methods, achieving state-of-the-art results.}, archive = {J_ASOC}, author = {Ruifeng Xu and Yi Chen and Zhongyan Yi and Shun Huang and Liang He}, doi = {10.1016/j.asoc.2025.113696}, journal = {Applied Soft Computing}, month = {12}, pages = {113696}, shortjournal = {Appl. Soft. Comput.}, title = {DTKD-DL: Dual-teacher knowledge distillation with dual-loops for continuous few-shot relation extraction}, volume = {184}, year = {2025}, }
• (2025). A two-stage feature selection approach with fuzzy covering-based rough sets based on discernibility matrix. ASOC, 184, 113695. (www)

  @article{ , abstract = {Fuzzy β -covering has attracted considerable research attention due to its enhanced capability to accurately and comprehensively represent uncertain information. Unlike partition-based approaches, fuzzy β -covering maintains its covering properties when augmented with additional elements. This inherent flexibility necessitates rigorous redundancy analysis. Hence, the calculation of the reduct of fuzzy β -covering relative to decision attribute constitutes a fundamental challenge in this context. To address this issue, we formally define the concepts of indispensable covering elements and core of fuzzy β -covering with respect to the decision attribute. Then, to facilitate efficient computation of core and reduct, we introduce the discernibility matrix for fuzzy β -covering and establish equivalent representations of the core and reduct. Furthermore, we conduct systematic examinations to verify the indispensability of some existing concepts in fuzzy β -covering. Meanwhile, a significant limitation of current feature selection methods useing fuzzy β -covering lies in their computational complexity, primarily due to repeated recalculation of dependency functions during iterations. To overcome this limitation, we propose an efficient feature selection algorithm that identifies all reducts directly through the discernibility matrix. Furthermore, we propose a two-stage feature selection method for fuzzy β -covering, which iteratively eliminates redundant features to identify optimal feature subsets in each iteration. Finally, we verify through time complexity analysis and numerical experiments that the proposed algorithm significantly outperforms existing feature selection algorithms in computational efficiency. Comparative analysis further reveals that our method achieves superior results in terms of selected feature subset sizes and classification accuracy.}, archive = {J_ASOC}, author = {Tianyu Wang and Shuai Liu and Bin Yang}, doi = {10.1016/j.asoc.2025.113695}, journal = {Applied Soft Computing}, month = {12}, pages = {113695}, shortjournal = {Appl. Soft. Comput.}, title = {A two-stage feature selection approach with fuzzy covering-based rough sets based on discernibility matrix}, volume = {184}, year = {2025}, }
• (2025). A multi-source remote sensing image watermarking model based on multi-domain generative adversarial networks. ASOC, 184, 113694. (www)

  @article{ , abstract = {Due to the heterogeneity stemming from distinct sensor types, quantized spatial resolution levels, and categorized acquisition angles within multi-source remote sensing images, their data distribution and characteristics vary significantly, leading to poor cross-domain consistency and weak resistance to domain-specific distortions in traditional watermarking approaches. Traditional watermarking methods often struggle to handle such heterogeneous datasets. To address this challenge, this paper presents a watermarking model based on a multi-domain generative adversarial network (MD-GAN). By incorporating a multi-domain adversarial training mechanism and designing multiple domain discriminators, the generator is optimized to work across various image domains, thereby enhancing model robustness. Additionally, a multi-scale generator is utilized to account for differences in spatial resolution, improving both the watermark’s concealment and extraction accuracy. Experimental results demonstrate that the MD-GAN model achieves a 7 % improvement in robustness over the state-of-the-art models in terms of resistance to noise attacks and geometric transformations. Ablation studies confirm that the multi-domain adversarial training mechanism and multi-scale generator contribute significantly to this enhancement. MD-GAN provides a powerful solution for the copyright protection of multi-source remote sensing images, with substantial potential for real-world applications such as secure transactions and data sharing.}, archive = {J_ASOC}, author = {Heyan Wang and Xingxiang Jiang and Minxuan Wang and Changqing Zhu and Na Ren and Luanyun Hu}, doi = {10.1016/j.asoc.2025.113694}, journal = {Applied Soft Computing}, month = {12}, pages = {113694}, shortjournal = {Appl. Soft. Comput.}, title = {A multi-source remote sensing image watermarking model based on multi-domain generative adversarial networks}, volume = {184}, year = {2025}, }
• (2025). Multiscale convolution-attention model for efficient alzheimer’s disease and mild cognitive impairment diagnosis approach using sMRI. ASOC, 184, 113693. (www)

  @article{ , abstract = {Alzheimer's disease (AD) affects over 50 million people worldwide and causes a gradual decline in memory, language, and actions, with no known cure. Early identification and treatment are crucial to improve the quality of life of affected individuals. Deep learning algorithms based on brain structural magnetic resonance imaging (sMRI) show promise in predicting AD. However, relying solely on deep convolutional neural network (CNN) architecture has limitations in global modeling. Transformer, which is a feature learning technique, has excelled in computer vision applications, sparking interest in its application in brain image processing. However, pure Transformer architectures encounter challenges when trained on small sMRI datasets. Meanwhile, CNN-based methods do not consider the interdependence between voxels, which hinders the comprehensive understanding of the global characteristics of sMRI data. To address these challenges, an effecient deep learning framework was developed that combines the benefits of both the CNN and Transformer. In the preliminary phase of feature extraction, we introduce a multiscale feature fusion stem that employs convolutional kernels of varying scales to derive low-level features from the sMRI and integrate them to enhance the recognition accuracy. Furthermore, the proposed method introduces convolutional split attention with a squeeze and excitation block and additional convolution operations in the core sections of the Transformer, thereby enabling multiscale feature extraction and fusion. The model integrates multi-head light self-attention and a sandglass local feed-forward network block for classifier modeling, enhancing the extraction of sMRI features with a MobileNet cost. The proposed model, which was tested on the Alzheimer's Disease Neuroimaging Initiative dataset, achieved superior performance in AD diagnosis with fewer parameters and reduced computational costs, demonstrating its potential as a state-of-the-art solution.}, archive = {J_ASOC}, author = {Uttam Khatri and Jun-Hyung Kim and Goo-Rak Kwon}, doi = {10.1016/j.asoc.2025.113693}, journal = {Applied Soft Computing}, month = {12}, pages = {113693}, shortjournal = {Appl. Soft. Comput.}, title = {Multiscale convolution-attention model for efficient alzheimer’s disease and mild cognitive impairment diagnosis approach using sMRI}, volume = {184}, year = {2025}, }
• (2025). F2CAU-net: A dual fuzzy medical image segmentation cascade method based on fuzzy feature learning. ASOC, 184, 113692. (www)

  @article{ , abstract = {Accurate medical image segmentation plays a pivotal role in clinical diagnosis and treatment planning. However, existing methods – particularly those based on U-Net architectures – still face considerable difficulties when dealing with ambiguous boundaries, low-contrast regions, and noise artifacts. These challenges arise from the deterministic nature of conventional convolutional and attention mechanisms, which are often inadequate in modeling the inherent uncertainty and variability present in medical images. To tackle these limitations, we propose F 2 CAU-Net, a Dual Fuzzy Medical Image Segmentation Cascade Method that integrates a fuzzy convolution module for better modeling of uncertain and imprecise features, and a fuzzy attention mechanism to suppress redundancy and enhance focus on regions of interest. The proposed method captures both local and global contextual fuzzy information to improve segmentation accuracy and robustness. Extensive experiments on multiple benchmark datasets – including COVID-19 lesions, brain tumors, and skin cancer – demonstrate that F 2 CAU-Net significantly outperforms existing state-of-the-art models, particularly in scenarios with complex boundaries and noise. This approach offers promising potential for clinical applications, providing a more reliable and uncertainty-aware solution for medical image analysis.}, archive = {J_ASOC}, author = {Tianyi Zhou and Haipeng Wang and Sheng Geng and Hengrong Ju and Jiashuang Huang and Fan Fu and Weiping Ding}, doi = {10.1016/j.asoc.2025.113692}, journal = {Applied Soft Computing}, month = {12}, pages = {113692}, shortjournal = {Appl. Soft. Comput.}, title = {F2CAU-net: A dual fuzzy medical image segmentation cascade method based on fuzzy feature learning}, volume = {184}, year = {2025}, }
• (2025). A physics-informed mitigation method for DC microgrids under cyber attacks. ASOC, 184, 113691. (www)

  @article{ , abstract = {Cyber attacks pose serious threats to DC microgrids, making effective mitigation strategies very critical. However, most existing mitigation schemes ignore the physical characteristics of microgrids, which may lead to inaccurate attack detection and affect the mitigation performance. To address this issue, a physics-informed mitigation method for cyber attacks on DC microgrids is proposed. A cyber–physical framework of the DC microgrid is established, and the impact of cyber attacks on the microgrid is analyzed. The denoising autoencoder is utilized to improve the quality of the input data, and then the consistency characteristic of the DC microgrid is incorporated as a physical-informed constraint into the training of the state estimation model. Subsequently, anomaly detection is performed by comparing the estimated state and the real-time measured state. Once the attacks are detected, the estimated state is used to compensate the control inputs, mitigating the effect of attacks and ensuring the safe and stable operation of the DC microgrid. Simulation results show that compared with existing neural network-based estimators, the proposed physics-informed state estimation model can enhance the accuracy by 7 . 98 % ∼ 12 . 54 % , providing more precise estimated states to mitigate the effects of attacks.}, archive = {J_ASOC}, author = {Wanwan Ren and Jun Peng and Yun Zhou and Weirong Liu and Fu Jiang}, doi = {10.1016/j.asoc.2025.113691}, journal = {Applied Soft Computing}, month = {12}, pages = {113691}, shortjournal = {Appl. Soft. Comput.}, title = {A physics-informed mitigation method for DC microgrids under cyber attacks}, volume = {184}, year = {2025}, }
• (2025). Acquisition of representative data sets by filtering out redundant objects and attributes with fuzzy preference-based rough sets and dominance principles. ASOC, 184, 113690. (www)

  @article{ , abstract = {The dominance principle is crucial for evaluating consistency in dominance-based rough sets, yet redundant objects or attributes impair decision consistency. While existing work focuses on modeling and attribute reduction, object-induced inconsistency remains understudied. This study proposes a novel methodology for acquiring compact datasets through representative object extraction and attribute reduction in fuzzy preference-based rough sets, with a specific focus on preserving consistency in dominance principles. Firstly, the extent of dominance relations is quantified by fuzzy preference relations, while the evaluation of consistency between conditional and decision attributes is accomplished through distance measures. Subsequently, representative objects are identified by eliminating those with lower consistency in dominance principles, as assessed by distance measures with a predefined parameter. Further, a streamlined dataset is achieved through attribute reductions in fuzzy preference-based rough sets, incorporating representative objects. In conclusion, our proposed method is validated using numerical datasets, and its effectiveness is evaluated through measures rooted in rough set theory, machine learning, and statistics. This research contributes to a more comprehensive understanding of dominance-based rough sets by addressing the often-overlooked issue of inconsistency resulting from redundant objects.}, archive = {J_ASOC}, author = {Shuyun Yang and Guang Shi and Yuchao Li}, doi = {10.1016/j.asoc.2025.113690}, journal = {Applied Soft Computing}, month = {12}, pages = {113690}, shortjournal = {Appl. Soft. Comput.}, title = {Acquisition of representative data sets by filtering out redundant objects and attributes with fuzzy preference-based rough sets and dominance principles}, volume = {184}, year = {2025}, }
• (2025). Spiking neural network with time-varying weights for rail squat detection. ASOC, 184, 113689. (www)

  @article{ , abstract = {Axle box acceleration (ABA) measurements can be used for continuously monitoring rail infrastructure and detecting rail surface defects such as squats. However, accurately detecting squats is challenging due to their short-duration responses and low occurrence in ABA signals, particularly for light squats that exhibit subtle ABA responses. To address this challenge, we propose using a spiking neural network (SNN) with time-varying weights to enhance the detection performance of rail squats based on ABA measurements. Our approach employs a simple SNN architecture without hidden layers, trained using a method that combines genetic algorithms, k-fold cross-validation, and multi-start gradient-based approach to optimise hyperparameters and weights. The proposed methodology demonstrates competitive accuracy compared to other state-of-the-art SNN-based methods on UCI benchmarks for both binary and multi-class nonlinear problems. Part of its advantages include higher efficiency with a simpler architecture and training approach that reduces computational times while achieving effective spatiotemporal pattern detection. As shown by real-field measurements from Dutch and Swedish railways in anomaly detection, it effectively captures subtle changes in light squat defect responses in ABA signals and achieves a detection performance of 100% for severe squat defects and over 93% for light squat defects. Furthermore, we show that the spike responses, postsynaptic potentials, and membrane potentials can be used as a new way to explain and analyse the ABA signals. The proposed method using time-varying weights highlights a correspondence with the physical problem and offers an ability to capture sudden and subtle changes in the responses, which is crucial, particularly for detecting defects in their early stages.}, archive = {J_ASOC}, author = {Wassamon Phusakulkajorn and Jurjen Hendriks and Zili Li and Alfredo Núñez}, doi = {10.1016/j.asoc.2025.113689}, journal = {Applied Soft Computing}, month = {12}, pages = {113689}, shortjournal = {Appl. Soft. Comput.}, title = {Spiking neural network with time-varying weights for rail squat detection}, volume = {184}, year = {2025}, }
• (2025). Spectral–spatial representation progressive learning via segmented attention for 3D skeleton-based motion prediction. ASOC, 184, 113688. (www)

  @article{ , abstract = {Recently, GCNs-based methods have demonstrated impressive performance in human behavior prediction tasks. We believe that human motion modeling can explained as motion correlation extraction from the combination of the active and static motion parts analysis. However, existing methodologies fail to address the issue that feature information associated with static regions may overshadow feature information from dynamic regions, ultimately affecting the extraction of network features. Moreover, the unique low-pass feature pre-retention processing mechanism of GCN on the spectrum will lead to the attitude of some sequences remaining unchanged during the prediction process and further hurt the prediction. In this paper, we propose a Spectral–Spatial Representation Progressive Learning network to solve the problem above. Firstly, we propose a segmented attention block to compare the input observation sequence with the static contrast standard to obtain the motion region and the rest region. Then, we design the Spectrum Deconstruction Recombination Factor block(SDRF) to extract the global bandpass spectrum of human bone joints. The joint features of different regions are encoded by graph convolution and high-frequency feature filter coding based on geometric algebra. Specifically, a spectral–spatial interaction block is presented in each SDRF, focusing on the diversity of motion sequence frequency domain and spatial domain map, and realizes the fine extraction of historical pose sequence features from the two levels of space and spectral domain. Experimental results demonstrate that our approach outperforms state-of-art algorithms by 2.4%, 5.3% and 4.7% in terms of 3D mean per joint position error on Human 3.6M, CMU Mocap and 3DPW datasets, respectively.}, archive = {J_ASOC}, author = {Wenming Cao and Jianhua Zhang and Jianqi Zhong}, doi = {10.1016/j.asoc.2025.113688}, journal = {Applied Soft Computing}, month = {12}, pages = {113688}, shortjournal = {Appl. Soft. Comput.}, title = {Spectral–spatial representation progressive learning via segmented attention for 3D skeleton-based motion prediction}, volume = {184}, year = {2025}, }
• (2025). Multi-modal multi-trip supply chain model aided by smart contract victim tracking — An innovative pathway to disaster management under uncertainty. ASOC, 184, 113687. (www)

  @article{ , abstract = {This study addresses the complex challenges of post-earthquake rescue operations, using the recent earthquake in Turkey as a reference case. It focuses on developing an optimized model for scenarios with acute vehicle shortages, aiming to minimize both operational costs and response time during the critical initial phase of disaster relief. The proposed solution is built upon a robust mathematical framework that employs aerial vehicles for post-disaster area assessment, resource allocation, and the relocation of critically injured victims. The model leverages a soft computing approach, integrating the Weighted Sum Method (WSM) and Neutrosophic Compromise Programming (NCP). To enhance decision-making under uncertainty, the framework incorporates hexagonal type-2 fuzzy defuzzification, a technique grounded in soft computing principles. Results demonstrate the effectiveness of this approach: the NCP method achieved a response time of 213 min (3.55 h) and a cost of Rs 821,026.5, compared to 217.5 min (3.62 h) and Rs 820,860.3 for the WSM method—both successfully coordinating the rescue of 1,450 victims through efficient deployment of drones and helicopters. In addition, the study introduces a decentralized Ethereum-based smart contract to securely store and retrieve critical victim information. Validated through rigorous unit testing, the contract ensures data transparency and integrity, executing at a cost of 0.00379246 Ether. This blockchain-enabled feature complements the core optimization model, supporting real-time, tamper-proof data handling. To further validate the model’s applicability, a second real-life numerical example based on the recent Sikkim cloudburst is analyzed. The findings reinforce the model’s adaptability and practical value. The managerial implications of this research highlight the importance of soft computing-driven decision support, proactive contingency planning, and the integration of intelligent technologies in disaster response. This holistic framework — combining soft computing methodologies, advanced optimization models, and blockchain technology — offers an innovative and scalable solution for enhancing the resilience and efficiency of disaster management supply chains.}, archive = {J_ASOC}, author = {Alisha Roushan and Amrit Das and Anirban Dutta and Uttam Kumar Bera}, doi = {10.1016/j.asoc.2025.113687}, journal = {Applied Soft Computing}, month = {12}, pages = {113687}, shortjournal = {Appl. Soft. Comput.}, title = {Multi-modal multi-trip supply chain model aided by smart contract victim tracking — An innovative pathway to disaster management under uncertainty}, volume = {184}, year = {2025}, }
• (2025). Contrastive learning with hard negatives for sentence embeddings. ASOC, 184, 113685. (www)

  @article{ , abstract = {Unsupervised sentence representation learning remains a core challenge in natural language processing. Recent contrastive learning methods have shown strong potential in capturing sentence-level semantics, but their effectiveness is often constrained by the quality of positive and negative samples. In particular, constructing informative hard negatives in textual data is significantly more difficult than in vision tasks due to the ambiguity and compositionality of natural language. We propose HNCSE, a novel unsupervised contrastive learning framework that enhances sentence representations through hard negative compositional strategies. HNCSE introduces two key components: HNCSE-HNM, which synthesizes informative hard negatives via mixup within the batch, and HNCSE-PM, which generates harder positives by leveraging the most challenging negatives. This joint design improves both alignment and discrimination in embedding space without relying on external supervision. Extensive experiments on semantic textual similarity and transfer tasks demonstrate that HNCSE consistently outperforms existing unsupervised and supervised baselines.}, archive = {J_ASOC}, author = {Wenxiao Liu and Zihong Yang and Chaozhuo Li and Zijin Hong and Jianfeng Ma and Zhiquan Liu and Litian Zhang and Feiran Huang}, doi = {10.1016/j.asoc.2025.113685}, journal = {Applied Soft Computing}, month = {12}, pages = {113685}, shortjournal = {Appl. Soft. Comput.}, title = {Contrastive learning with hard negatives for sentence embeddings}, volume = {184}, year = {2025}, }
• (2025). HumanMoD: A multi-RAG collaborative LLM for inclusive urban public healthcare services. ASOC, 184, 113684. (www)

  @article{ , abstract = {The integration of Large Language Models (LLMs) with healthcare systems offers transformative solutions for sustainable public health programs, especially in underserved urban and rural communities where access to professional medical expertise is limited. This paper introduces HumanMoD, a novel LLM framework designed to emulate collaborative clinical workflows of multi-expert medical doctors. Specifically, the proposed Mixture of Doctors module enables parallel diagnostic streams from diverse medical specialties, mimicking the collaborative decision-making of human doctors to ensure comprehensive assessments in resource-constrained environments. The Knowledge-driven Medical Assistant leverages domain-specific knowledge bases to mitigate LLM hallucinations, ensuring that recommendations are rooted in credible medical knowledge. Exquisitely, the Humanoid Health Conductor and LLM-powered Corrector further refine outputs to minimize diagnostic discrepancies, enhancing the reliability of responses for large-scale public health applications such as community health kiosks and mobile health apps serving remote or low-income areas. Unlike fine-tuned models, HumanMoD operates without parameter adjustment, enabling cost-effective deployment in regions with scarce computational resources, thus bridging the healthcare gap for socially vulnerable groups. Experimental results on MedQA and PubMedQA demonstrate that HumanMoD outperforms state-of-the-art models, highlighting its potential to drive equitable healthcare access and support data-driven public health policies in diverse urban and rural settings.}, archive = {J_ASOC}, author = {Song Sun and Zhijie Zhong and Nanlan Yu and Xinrong Gong and Kaixiang Yang}, doi = {10.1016/j.asoc.2025.113684}, journal = {Applied Soft Computing}, month = {12}, pages = {113684}, shortjournal = {Appl. Soft. Comput.}, title = {HumanMoD: A multi-RAG collaborative LLM for inclusive urban public healthcare services}, volume = {184}, year = {2025}, }
• (2025). A KPI-related fault diagnosis method for multimode manufacturing processes based on supervised minimal gated unit and sparse broad learning system. ASOC, 184, 113679. (www)

  @article{ , abstract = {In multimode manufacturing processes, the key performance indicator (KPI)-related fault diagnosis plays a critical role in ensuring product quality and enhancing economic benefits. However, variability in working modes, lack dynamic description for the KPIs, and model structure redundancy may lead to poor universality and low diagnostic accuracy of conventional deep learning-based approaches. In this work, a supervised minimal gated unit and sparse broad learning system (SMGU-SBLS) is developed for KPI-related fault diagnosis. Specifically, the KPIs and process variables are simultaneously utilized in the SMGU to learn the KPI-related dynamic features. Then, aiming at simplifying the network structure, a sparse version of broad learning system is proposed for fault diagnosis. Furthermore, the expansion capability of SMGU-SBLS has been analyzed. Finally, the proposed SMGU-SBLS network is applied to a real hot strip mill process (HSMP). Simulation results show that the proposed method has higher diagnostic performance than the other four state-of-the-art deep learning methods.}, archive = {J_ASOC}, author = {Chuanfang Zhang and Wenxiao Yin and Chi Zhang and Kaixiang Peng and Xueyi Zhang}, doi = {10.1016/j.asoc.2025.113679}, journal = {Applied Soft Computing}, month = {12}, pages = {113679}, shortjournal = {Appl. Soft. Comput.}, title = {A KPI-related fault diagnosis method for multimode manufacturing processes based on supervised minimal gated unit and sparse broad learning system}, volume = {184}, year = {2025}, }
• (2025). Green maritime transport performance analysis of G-7 countries using an interval-valued fermatean fuzzy ARLON-based decision model. ASOC, 184, 113677. (www)

  @article{ , abstract = {Maritime transport operations are witnessing heightened environmental sensitivities, prompting the implementation of environmentally focused maritime transport practices. The principal aim is to develop a decision support system for evaluating countries' green maritime transport performance. To this end, the interval-valued Fermatean fuzzy (IVFF)-simple weight calculation (SIWEC)-skewness impact through distributional evaluation (SITDE)-alternative ranking using two-step logarithmic normalization (ARLON) model is developed within this study. This method enables the analysis of countries' green maritime transport performance by integrating expert opinions with environmental and maritime transport parameters. The model determines the influence of experts using IVFF sets. It further facilitates the simultaneous use of subjective and objective criteria weighting approaches to compute the weights of green maritime transport performance criteria. ARLON is employed to assess and rank countries' green maritime transport performance levels. The applicability of the IVFF-SIWEC-SITDE-ARLON model is tested through a case study focusing on G-7 countries, and the results supported the successful implementation of the method. Furthermore, sensitivity and comparative analyses demonstrated the consistency and robustness of the model. According to the findings of the case study, the United States emerges as the country with the highest green maritime transport performance among the G-7 countries. The "linear shipping connectivity index" is identified as the most influential criterion in the decision-making process. The study offers actionable recommendations for the maritime industry, thereby contributing to the advancement of green maritime transport practices.}, archive = {J_ASOC}, author = {Galip Cihan Yalçın and Karahan Kara and Emre Kadir Özekenci and Vladimir Simic and Dragan Pamucar}, doi = {10.1016/j.asoc.2025.113677}, journal = {Applied Soft Computing}, month = {12}, pages = {113677}, shortjournal = {Appl. Soft. Comput.}, title = {Green maritime transport performance analysis of G-7 countries using an interval-valued fermatean fuzzy ARLON-based decision model}, volume = {184}, year = {2025}, }
• (2025). Adaptive self-correction network for human motion prediction. ASOC, 184, 113676. (www)

  @article{ , abstract = {Human motion prediction aims to generate future poses from the observed historical human motion sequence. It is fundamental to many intelligent systems, e.g., human–robot interaction and self-driving. Though the existing encoder–decoder methods obtain good performance in some scenarios, there is still a gap between their prediction results and ground truth in many cases. In this work, we propose to estimate the deviation between the decoding results (which will be referred to as the conventional human motion prediction ) and the groundtruth, and integrate this estimation with the conventional prediction results to derive the corrected human motion prediction . In this way, our method can self-correct the conventional prediction results based on a preliminary estimated deviation from it to the groundtruth, and thus enhance the performance. In our work, we adopt five independent lightweight branches rather than a global estimator to estimate the deviation of the five human body components ( i.e., left arm, right arm, torso, left leg, and right leg). Based on this component-wise deviation estimation strategy, we propose a Fixed Self-Correction Network (FSCNet) for human motion prediction to obtain enhanced performance. Recognizing that not all joints exhibit the same motion dynamics inside one given body component, we further propose the Adaptive Self-Correction Network (ASCNet) to let these five estimators adaptively capture the correlated deviations and thus enhance human motion prediction performance. Extensive experiments on three large datasets (Human3.6M, CMU-Mocap, and 3DPW) validate the superiority of our proposed FSCNet and ASCNet over the established works.}, archive = {J_ASOC}, author = {Jinkai Li and Jinghua Wang and Xin Wang and Liang Yan and Xiaoling Luo and Yong Xu}, doi = {10.1016/j.asoc.2025.113676}, journal = {Applied Soft Computing}, month = {12}, pages = {113676}, shortjournal = {Appl. Soft. Comput.}, title = {Adaptive self-correction network for human motion prediction}, volume = {184}, year = {2025}, }
• (2025). RGB-D indoor scene parsing via wavelet sub-band guided transformer. ASOC, 184, 113675. (www)

  @article{ , abstract = {Depth information has been shown to be complementary to RGB scene parsing. However, inherent disparities between RGB and depth modalities create challenges for effective feature fusion. Furthermore, current methods often lose high-frequency information during downsampling, limiting the full utilization of RGB and depth image information. To address these issues, we propose the Wavelet Sub-band Guided Transformer (WSGFormer), which utilizes wavelet sub-band to enhance feature correction, fusion, and refinement. The WSGFormer contains three important modules. Firstly, the Wavelet Cross-attention Rectification Module employs Haar wavelet transforms to decompose features into wavelet sub-band, and adaptively aligns RGB and depth features by extracting their mapping relationships. Secondly, the Multi-scale Fusion Module combines RGB and depth branches, utilizing vertical bar-shaped convolutions to enable cross-modal feature selection and enhance the sensitivity to high-frequency components through frequency-aware techniques. Finally, the Discrepancy Compensation Module starts with high-level semantic information and progressively guides the fusion of adjacent layers downwards, reducing disparities between them through subtraction operations. The evaluation conducted on the NYUv2, SUN-RGBD and ScanNetV2 datasets highlights the superior performance of the proposed WSGFormer.}, archive = {J_ASOC}, author = {Wen Xie and Heng Liu and JiaHao Li}, doi = {10.1016/j.asoc.2025.113675}, journal = {Applied Soft Computing}, month = {12}, pages = {113675}, shortjournal = {Appl. Soft. Comput.}, title = {RGB-D indoor scene parsing via wavelet sub-band guided transformer}, volume = {184}, year = {2025}, }
• (2025). Large language model driven multiple operating conditions identification and predictive control. ASOC, 184, 113674. (www)

  @article{ , abstract = {Dynamic changes in operating conditions are common in industrial process control. Efficient and accurate detection of condition transitions and identification of operating states are critical for achieving precise control across multiple operating conditions. Traditional condition identification methods primarily rely on the numerical similarity of input–output sequences, often neglecting the dynamic semantic information embedded within the data, such as variation rates and overshoot magnitudes, which can lead to frequent misclassifications. Recently, large language models (LLMs) have exhibited remarkable capabilities in semantic understanding of complex sequences, offering a new perspective for the identification of operating conditions. However, their application in industrial control still faces two major challenges. First, the enormous parameter scale of LLMs results in high computational costs, making it difficult to achieve real time comparisons between online sequences and large historical datasets, thereby compromising the timeliness of condition identification. Second, LLMs are designed primarily for textual input, leading to a significant modality gap when processing numerical input–output sequence data, which limits their full semantic understanding potential in industrial scenarios. To address these challenges, this paper proposes a novel large language model driven multiple operating conditions identification and predictive control method. The proposed method fully leverages the semantic understanding capabilities of LLMs by mining the underlying dynamic characteristics of input–output sequences, enabling rapid identification of operating conditions under limited sample scenarios and achieving precise control across multiple conditions. Specifically, first, to accurately detect condition changes in multiple operating condition industrial processes, an Approximate Entropy based operating condition change detection method is proposed. Considering that condition transitions often cause a prediction model mismatch, leading to increased complexity and irregularity in control sequences, approximate entropy is employed to quantify the sequence complexity. A data driven adaptive thresholding mechanism based on kernel density estimation is further developed to achieve robust detection of condition changes. Second, to address the issues of limited samples and low identification accuracy arising from solely relying on numerical features, a novel LLM driven multimodal condition identification method is proposed. This method constructs a sparse representation based prediction model for historical operating conditions, forming a high information density knowledge base consisting of prediction models and representative sequences. To bridge the modality gap between numerical and textual data, a numerical to textual sequence description method enriched with dynamic semantics is innovatively introduced, enabling effective alignment between the two modalities. Furthermore, a dynamic semantics enhanced prompt engineering strategy is developed to fully exploit the LLM’s semantic understanding capabilities, facilitating accurate condition identification even under limited sample conditions. Finally, to achieve rapid response and precise control following condition changes, the proposed method quickly matches a suitable prediction model based on the LLM driven condition identification results and dynamically adjusts control parameters using a rolling horizon optimization strategy, thereby significantly improving control accuracy across multiple operating conditions. Notably, the proposed approach eliminates the need for reconstructing condition models after changes, ensuring smooth and continuous control under dynamic conditions. Extensive experiments verified the superiority of the proposed method in terms of both condition identification and multiple operating conditions control performance.}, archive = {J_ASOC}, author = {Zhongyu Zhang and Minzhi Mao and Keke Huang and Dehao Wu and Chunhua Yang}, doi = {10.1016/j.asoc.2025.113674}, journal = {Applied Soft Computing}, month = {12}, pages = {113674}, shortjournal = {Appl. Soft. Comput.}, title = {Large language model driven multiple operating conditions identification and predictive control}, volume = {184}, year = {2025}, }
• (2025). MetaIndux-PLC: A control logic-guided LLM for PLC code generation in industrial control systems. ASOC, 184, 113673. (www)

  @article{ , abstract = {Programmable Logic Controllers (PLCs) are widely used for automation control, and they are well-suited for industrial systems control tasks. LLMs can assist engineers in streamlining the programming process and reducing development costs, and one of the key issues is the construction of the PLC code dataset. However, the lack of an open-source PLC code dataset in the research community, combined with the high complexity of industrial systems control logic, has caused most LLMs to struggle with generating accurate control code. This complexity arises from the need to manage real-time sensor data fusion, integrate various communication protocols, and ensure compliance with stringent safety and regulatory standards. In this study, we construct ST4Indux, a PLC code dataset specifically for industrial systems control. And we propose the Control Logic-Guided Iterative Fine-Tuning (CLIFT) method, which iteratively optimizes the model’s generation capability. Based on these, we train a large language model named MetaIndux-PLC, to enable the generation of complex motion control code. Additionally, we propose a multi-dimensional evaluation and optimization method to systematically assess the model’s performance in terms of task completion quality, efficiency, and user experience. The experimental results demonstrate that the proposed approach significantly enhances MetaIndux-PLC’s performance and reliability in real-world engineering environments, providing a foundation for the future development of intelligent programming assistance systems.}, archive = {J_ASOC}, author = {Lei Ren and Haotian Wang and Jiabao Dong and Haiteng Wang and Shuai Liu and Yuanjun Laili and Lin Zhang}, doi = {10.1016/j.asoc.2025.113673}, journal = {Applied Soft Computing}, month = {12}, pages = {113673}, shortjournal = {Appl. Soft. Comput.}, title = {MetaIndux-PLC: A control logic-guided LLM for PLC code generation in industrial control systems}, volume = {184}, year = {2025}, }
• (2025). Two-stage identification of false data injection attacks in power systems via semi-supervised deep learning. ASOC, 184, 113672. (www)

  @article{ , abstract = {Previous studies find that malicious data manipulations against the state estimation, such as the false data injection attack (FDIA), can evade the detection of a conventional bad data detector equipped with a measurement meter. This calls for advanced FDIA identification approaches urgently. However, existing data-driven efforts are designed under the assumption that attacks are frequent and the amount of compromised data is comparable to benign data, which may not be realistic. Thus, these approaches deliver unsatisfactory performance under highly imbalanced data in the real world. To overcome this issue, we propose a novel two-stage FDIA identification pipeline, which formulates the problem as global detection and fine-grained localization. Following this framework, we leverage deep support vector data description to distinguish attacks from benign measurements in an unsupervised manner and employ a modified one-dimensional ResNet to locate the attacking aims upon detecting an FDIA. Our approach can overcome existing limitations induced by data-driven methods under infrequent FDIAs, leading to effective and robust FDIA identification. Case studies on IEEE standard 14-bus and 118-bus systems demonstrate the effectiveness and superiority of our approach and validate our findings.}, archive = {J_ASOC}, author = {Fengrui Liu and Keng-Weng Lao and Yida Xu and Yang Li and Haotian Guo and Xiaorui Hu and Yikun Yin}, doi = {10.1016/j.asoc.2025.113672}, journal = {Applied Soft Computing}, month = {12}, pages = {113672}, shortjournal = {Appl. Soft. Comput.}, title = {Two-stage identification of false data injection attacks in power systems via semi-supervised deep learning}, volume = {184}, year = {2025}, }
• (2025). Dynamic target continuous assignment method for unmanned clusters in fragmented information environments. ASOC, 184, 113671. (www)

  @article{ , abstract = {Target assignment stands as a pivotal resource management technique in collaborative clustering. During far-sea search and rescue (SAR) missions, challenges arise in formation control under fragmented information, stemming from unstable communication and incomplete data. Additionally, the dynamic nature of targets poses continuous assignment challenges. Addressing these issues in cluster collaboration represents a significant challenge. This paper proposes a method for dynamic target successive allocation tailored to fragmented information environments. First, inspired by biological cluster behaviors and integrating graph theory and complex network theory, the approach achieves decentered cluster formation by managing subgroup separation and aggregation. This solution effectively addresses the formation control challenge in fragmented information settings during cluster collaboration. Second, leveraging reinforcement learning principles, the method determines device behavioral strategies based on maximizing the device’s interaction rewards with the environment. This approach resolves the dynamic continuous target assignment challenge within clustered environments. This paper proposes a Broken Info-Driven Target Assignment (BI-DTA) method to address these research challenges. Experimental results demonstrate that the method achieves effective decentered formation control and dynamic continuous target assignment rapidly, exhibiting robustness and stability in fragmented information environments.}, archive = {J_ASOC}, author = {Rui Ding and Yuhan Zhu and Baojie Chai}, doi = {10.1016/j.asoc.2025.113671}, journal = {Applied Soft Computing}, month = {12}, pages = {113671}, shortjournal = {Appl. Soft. Comput.}, title = {Dynamic target continuous assignment method for unmanned clusters in fragmented information environments}, volume = {184}, year = {2025}, }
• (2025). A surrogate-assisted evolutionary algorithm based on the two-round selection strategy incorporating local search for expensive high-dimensional multi/many-objective optimization. ASOC, 184, 113670. (www)

  @article{ , abstract = {Surrogate-assisted multi-objective optimization has exhibited excellent performance for solving optimization problems that involve time-intensive computer simulations or resource-intensive physical experiments. However, the majority of existing research has focused on low-dimensional problems. In this paper, a surrogate-assisted evolutionary algorithm based on the two-round selection strategy incorporating local search (TRLS) is proposed for expensive high-dimensional multi/many-objective optimization. Specifically, the convergence and diversity of the trial solutions are assessed based on the estimation of Pareto fronts, and the uncertainty is quantified by analyzing the distribution of relevant points in the decision space. The infill sampling task is guided by the above three performance indicators. Firstly, a preliminary screening is conducted by considering convergence and diversity performance. Subsequently, the quality of candidates is further refined through the implementation of a local search strategy. Finally, a comprehensive fitness is constructed to select the sampling individual. In addition, a dynamic termination criterion is devised for the surrogate-assisted evolution phase. Empirical studies, conducted on two classical benchmark suites and two real-world tasks, reveal the effectiveness and applicability of the proposed TRLS.}, archive = {J_ASOC}, author = {Yang Li and Weigang Li and Songtao Li and Qifeng Wang and Junwei Hu}, doi = {10.1016/j.asoc.2025.113670}, journal = {Applied Soft Computing}, month = {12}, pages = {113670}, shortjournal = {Appl. Soft. Comput.}, title = {A surrogate-assisted evolutionary algorithm based on the two-round selection strategy incorporating local search for expensive high-dimensional multi/many-objective optimization}, volume = {184}, year = {2025}, }
• (2025). A study of the hesitant fuzzy aggregation operators for the transportation of perishable goods under real life scenarios. ASOC, 184, 113668. (www)

  @article{ , abstract = {The transportation of the perishable goods has always been a critical challenge in the market. The inherent susceptibility of the perishable items to spoilage and deterioration during transit necessitates the development and implementation of the preservation technologies that enhance the shelf life and maintain the quality of perishable products during transportation. Also, the carbon emissions stemming from activities associated with transportation present a significant and progressively urgent challenge, markedly contributing to environmental degradation. Consequently, the escalating carbon emissions linked to the transportation-related activities pose a threat to the sustainability of the environment. Recalling these facts, this paper explores the study of an interval valued multi-objective fixed charge solid transportation problem under hesitant fuzzy aggregation operators. Furthermore, this paper introduces a practical mathematical framework designed to represent the decision-making process inherited in the transportation scenarios. The proposed method leverages the hesitant fuzzy aggregation operators to offer a specific approach for the decision-making across such operators. Additionally, it introduces the notion of hesitant degrees for different objectives through the utilization of membership functions. Three conflicting objective functions: time minimization for customer satisfaction, profit maximization for the economic sustainability and minimization of the carbon emissions for the environmental sustainability have been addressed in the suggested model. The transportation time and the fuel consumption have been managed by introducing a variable called the vehicle speed coefficient. An analysis of the deterioration rates of the perishable products has been conducted considering the presence of a preservation value. Moreover, the proposed model has been formulated by employing the diverse approaches and have been solved using multi-objective techniques. A real-life-based numerical problem is presented and solved to validate the proposed concept. The results are compared with respect to different vehicle speeds and preservation values.}, archive = {J_ASOC}, author = {Awdhesh Kumar Bind and Deepika Rani and Ali Ebrahimnejad}, doi = {10.1016/j.asoc.2025.113668}, journal = {Applied Soft Computing}, month = {12}, pages = {113668}, shortjournal = {Appl. Soft. Comput.}, title = {A study of the hesitant fuzzy aggregation operators for the transportation of perishable goods under real life scenarios}, volume = {184}, year = {2025}, }
• (2025). SAGE-net: Single-layer augmented gated encoder network for efficient multimodal sentiment analysis. ASOC, 184, 113665. (www)

  @article{ , abstract = {Multimodal sentiment analysis has achieved remarkable progress. However, the increasing computational complexity of existing models poses significant challenges in resource-constrained scenarios. To address these challenges, this study introduces a single-layer augmented gated encoder network (SAGE-Net), a novel lightweight multimodal sentiment analysis architecture. In contrast to conventional multilayer, deeply stacked architectures, SAGE-Net employs only a single-layer encoder to preserve multimodal feature comprehension, significantly reducing computational complexity. To enable effective inter-modal interaction, a single-layer cross-attention mechanism is integrated. We extensively evaluate multiple feature fusion strategies and data augmentation strategies to enhance model effectiveness. Experimental results on the CMU-MOSI and CMU-MOSEI, and CH-SIMS datasets demonstrate that SAGE-Net achieves competitive performance and significantly lowers model size and tuning costs. These results validate SAGE-Net as a viable lightweight solution for multimodal sentiment analysis in resource-constrained scenarios.}, archive = {J_ASOC}, author = {Jiazheng Zhou and Xin Kang and Weiping Ding and Linhuang Wang and Fei Ding and Kazuyuki Matsumoto and Chenmeng Zhang and Huiwen Chi}, doi = {10.1016/j.asoc.2025.113665}, journal = {Applied Soft Computing}, month = {12}, pages = {113665}, shortjournal = {Appl. Soft. Comput.}, title = {SAGE-net: Single-layer augmented gated encoder network for efficient multimodal sentiment analysis}, volume = {184}, year = {2025}, }
• (2025). Robust weakly supervised product surface defect segmentation based on guided cropping and inpainting extension. ASOC, 184, 113661. (www)

  @article{ , abstract = {In the realm of industrial manufacturing, automatic detection and segmentation of surface defects in products is vital to enhance both product quality and efficiency. However, a large number of existing deep learning methods require a substantial amount of manually labeled data for training, and the high-cost labeling process hampers the practical application of such methods. Towards this end, we present a weakly supervised defect segmentation algorithm without any segment labels. First, a training enhancement method based on a contrastive learning module (CLM) coupled with a guided cropping module (GCM) is proposed to improve the network’s attention to defects in the Defect Focus Classifier (DFC) training phase. Subsequently, a novel inpainting extension module (IEM) generates a final class activation map (CAM) to obtain a pseudo label automatically for segment network training. Finally, conditional random field (CRF) and an additional training round refine the segmentation results. Moreover, the whole process is distilled into a fully supervised segmentation network to improve the inference efficiency. Conducting extensive experimentation, we have achieved 100% and 93.39% average precision (AP) and 41.73% and 53.14% average intersection-over-union (IOU) on the public datasets KolektorSDD and KolektorSDD2, respectively. Furthermore, we verified the generalizability of our method by conducting experiments on several industrial product classes in the MVTec AD, MTD, and DAGM datasets. In these experiments, we achieved favorable classification and segmentation results using solely image classification labels.}, archive = {J_ASOC}, author = {Rui Yan and Xiaojun Wu and Qixun Yang and Michael Yu Wang}, doi = {10.1016/j.asoc.2025.113661}, journal = {Applied Soft Computing}, month = {12}, pages = {113661}, shortjournal = {Appl. Soft. Comput.}, title = {Robust weakly supervised product surface defect segmentation based on guided cropping and inpainting extension}, volume = {184}, year = {2025}, }
• (2025). Semi-supervised lip-tongue segmentation with boundary region contrast sampling. ASOC, 184, 113653. (www)

  @article{ , abstract = {In Traditional Chinese Medicine, the accurate segmentation mask of the tongue and lip is the key of inspection. Although deep learning has made remarkable progress in medical image segmentation, a lot of manual annotations are still required for training. Semi-supervised learning (SSL) is used to reduce annotation work, but its performance often suffers when applied to tongue and lip segmentation, which is because tongue and lip images have noisy background information and unique boundary regions. To alleviate the problem, we propose a semi-supervised framework named Lip-Tongue segmentation with Boundary Region Contrast Sampling (Lip-Tongue-BReCoSample). We first preprocess the data, roughly locating the target and filtering out noisy background information. Then we generate the key boundary regions and sample to carry out contrast learning, which alleviates the problem that SSL cannot make fine modeling of the boundary regions of the target with limited information. After a lot of experiments, our method has achieved good results in SSL, and makes it reach or even exceed the performance of many traditional supervised methods, which can improve MIOU performance from 88.09 to 90.43 (+2.34) in SSL specifically. Our method is also better than the latest large-dataset pre-trained model (e.g., SegGPT). To the best of our knowledge, it is the first application of SSL in tongue and lip semantic segmentation.}, archive = {J_ASOC}, author = {Tao Jiang and Lechao Zhang and Wang Yuan and Liping Tu and Ji Cui and Xiaojuan Hu and Xin Tan and Lizhuang Ma and Jiatuo Xu}, doi = {10.1016/j.asoc.2025.113653}, journal = {Applied Soft Computing}, month = {12}, pages = {113653}, shortjournal = {Appl. Soft. Comput.}, title = {Semi-supervised lip-tongue segmentation with boundary region contrast sampling}, volume = {184}, year = {2025}, }
• (2025). Short-term multistep wind direction prediction of unmanned sailboats based on OVMD and optimized deep learning model. ASOC, 184, 113651. (www)

  @article{ , abstract = {Accurate short-term multistep wind direction prediction is crucial for enhancing the sailing performance and operational safety of unmanned sailboats. Existing methods often face challenges due to the complexity, non-stationarity, and diverse frequency characteristics of wind direction data. In this study, a novel approach is proposed that combines optimal variational mode decomposition (OVMD) with optimized deep learning model for wind direction prediction. First, OVMD is applied to decompose the wind data into stable modal signals, effectively reducing the impact of data complexity and non-stationarity on prediction performance. Considering that different subsequences exhibit distinct frequency patterns, five deep learning models, including Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, are employed to predict each subsequence separately. The most suitable model for each subsequence is selected based on the root mean square error (RMSE) metric. Additionally, hyperparameter optimization is conducted to enhance prediction accuracy, reduce training time, and eliminate the need for subjective parameter settings. Experimental results demonstrate that the proposed method can accurately capture wind direction variations and achieves superior performance compared to baseline models across all evaluation metrics, ensuring high accuracy and stability in short-term multistep wind direction prediction.}, archive = {J_ASOC}, author = {Zhipeng Shen and Shaoqing Zhang and Yang Yang and Zhaoyang Wu and Haomiao Yu}, doi = {10.1016/j.asoc.2025.113651}, journal = {Applied Soft Computing}, month = {12}, pages = {113651}, shortjournal = {Appl. Soft. Comput.}, title = {Short-term multistep wind direction prediction of unmanned sailboats based on OVMD and optimized deep learning model}, volume = {184}, year = {2025}, }
• (2025). Attention-inspired text-based recommender system with explanatory capabilities. ASOC, 184, 113650. (www)

  @article{ , abstract = {Recommender systems are widely used to help users find relevant and personalized items in various domains. However, providing accurate recommendations is not enough to ensure user satisfaction, trust, and engagement. Nowadays, users demand transparency from these systems typically in the form of an explanation of the recommendation given. This paper presents a novel explainable Recommender System designed to generate recommendations from natural language queries while providing model-intrinsic explanations inspired by attention mechanisms. The system adds transparency, interpretability, and new user cold-start capabilities. We evaluate our approach on twelve datasets from diverse domains and languages, demonstrating its effectiveness and robustness. Results show that our proposal achieves competitive accuracy with respect to strong baselines, while consistently outperforming a prior interpretable model developed for the same task.}, archive = {J_ASOC}, author = {Pablo Pérez-Núñez and Paul Buitelaar and Jorge Díez and Oscar Luaces and Antonio Bahamonde}, doi = {10.1016/j.asoc.2025.113650}, journal = {Applied Soft Computing}, month = {12}, pages = {113650}, shortjournal = {Appl. Soft. Comput.}, title = {Attention-inspired text-based recommender system with explanatory capabilities}, volume = {184}, year = {2025}, }
• (2025). MHGCL: A multi-modal hypergraph contrastive learning framework for molecular property prediction. ASOC, 184, 113645. (www)

  @article{ , abstract = {Accurate molecular property prediction (MPP) is pivotal in drug discovery. Although current models integrate multi-modal (1D, 2D, and 3D) molecular features, they often suboptimally leverage pharmacophoric information and inadequately capture higher-order intramolecular relationships, such as conjugated systems. This study introduces a novel multi-modal hypergraph contrastive learning framework (MHGCL) to generate enriched molecular representations for enhanced MPP. MHGCL uniquely employs hypergraphs to model complex, many-to-many interactions within molecules. It incorporates a dual-channel architecture, featuring a hypergraph transformer and an equivariant graph neural network, to distinctly process 2D and 3D molecular information. Crucially, functional group and chemical element-oriented knowledge graphs are integrated to explicitly imbue the model with pharmacophoric knowledge. The contrastive learning strategy effectively aligns these diverse representations. Extensive experiments across ten benchmark datasets demonstrate that MHGCL consistently outperforms existing state-of-the-art methods. Further ablation studies confirm that the proposed hypergraph-based molecular representation captures structural motifs of molecular functional groups more effectively, thereby affirming the design efficacy of the model’s constituent modules.}, archive = {J_ASOC}, author = {Rui Han and Qun Liu and Xu Gong and Guoyin Wang and Li Liu and Xingping Xian}, doi = {10.1016/j.asoc.2025.113645}, journal = {Applied Soft Computing}, month = {12}, pages = {113645}, shortjournal = {Appl. Soft. Comput.}, title = {MHGCL: A multi-modal hypergraph contrastive learning framework for molecular property prediction}, volume = {184}, year = {2025}, }
• (2025). A multi-hop shapley-based framework for graph convolutional network node classification explanation. ASOC, 184, 113615. (www)

  @article{ , abstract = {Graph Convolutional Networks (GCNs) have recently demonstrated superior performance across various graph machine learning tasks. However, they are often regarded as “black-box” models, which has sparked significant interest in developing methods to interpret their predictions. Among various GCN explanation techniques, game-theoretic Shapley value approaches stand out for their ability to identify important nodes, edges, and features. Nonetheless, most Shapley-based explanation models tend to concentrate on dependencies within a fixed radius, resulting in a constrained perceptual field. Additionally, existing explanation methods primarily focus on the total marginal contributions of either very small or very large coalitions during sampling, which limits their effectiveness in capturing the joint marginal contributions inherent in mid-sized subgraphs. To address these challenges, we propose a novel Shapley-based GCN explanation model called MixHopShap, which incorporates multi-hop information and a balanced sampling strategy. MixHopShap employs a multi-hop computational graph construction process to generate an explanation domain, enabling it to capture both local and long-range dependencies. Moreover, MixHopShap introduces a new sampling strategy that promotes balanced coalition coverage, allowing for efficient sampling of mid-sized subgraphs and facilitating learning of marginal contributions for each edge. We conduct experiments on six real world datasets to evaluate the performance of MixHopShap in terms of Fidelity and ROAR metrics. The experimental results demonstrate the superiority of MixHopShap over state-of-the-art methods. Additionally, the explainability of MixHopShap can significantly improve the confidence of GCN’s prediction in node classification tasks.}, archive = {J_ASOC}, author = {Yifan Zheng and Xibei Yang and Qiguo Sun and Keyu Liu and Qihang Guo}, doi = {10.1016/j.asoc.2025.113615}, journal = {Applied Soft Computing}, month = {12}, pages = {113615}, shortjournal = {Appl. Soft. Comput.}, title = {A multi-hop shapley-based framework for graph convolutional network node classification explanation}, volume = {184}, year = {2025}, }
• (2025). Graph neural network-based interactive clustering enhanced by human knowledge. ASOC, 184, 113595. (www)

  @article{ , abstract = {Clustering techniques face persistent challenges in balancing automation with human interpretability. Traditional methods require laborious parameter tuning and domain expertise to define similarity measures and validate results, while deep learning approaches trade transparency for performance. To bridge this gap, we propose a human-in-the-loop framework that synergizes domain knowledge with graph-based semi-supervised learning. Our system enables users to iteratively refine clusters through intuitive visual adjustments on a subset of data, guided by real-time quality metrics to reduce errors and decision fatigue. These sparse annotations propagate to unlabeled instances via a graph neural network (GNN) that models latent relationships through modularity-driven structural learning. By translating cluster adjustments into semi-supervised classification tasks, our method eliminates manual feature engineering and scales to large datasets without retraining. Evaluations on two subsets of the MNIST dataset demonstrated that the NMI (Normalized Mutual Information) of our method improved by 50.44% and 64.77% relative to baseline clustering method, respectively.}, archive = {J_ASOC}, author = {Yunzhe Wang and Yushi Li and Qiming Fu and Chengtao Ji and You Lu and Jianping Chen}, doi = {10.1016/j.asoc.2025.113595}, journal = {Applied Soft Computing}, month = {12}, pages = {113595}, shortjournal = {Appl. Soft. Comput.}, title = {Graph neural network-based interactive clustering enhanced by human knowledge}, volume = {184}, year = {2025}, }
• (2025). Advanced deep learning approaches for fault detection in solar PV systems: A comparative study of SPDA and AIFD-SolDL. ASOC, 184, 113592. (www)

  @article{ , abstract = {Efficient operation of solar photovoltaic (PV) systems is critical for maximizing power generation and ensuring optimal energy conversion. However, faults in PV modules can significantly impact system performance and reduce energy output. Therefore, accurate identification and diagnosis of these malfunctions are essential. To address the challenge of fault detection in solar PV systems, this study presents two distinct approaches. The first, called Solar Panel Degradation Assessment (SPDA), evaluates faults in solar panels by analyzing degradation effects while considering environmental factors like radiation and temperature. The second approach, named Efficient Ensemble Deep Learning Model for Enhancing Fault Detection in Solar PV Systems (AIFD-SolDL), utilizes advanced deep learning techniques, including DenseNet201, Inception-ResNet-v2, and Inception-v3, to enhance fault detection accuracy. In the AIFD-SolDL approach, PV module data undergo deep feature extraction followed by dimensionality reduction using principal component analysis (PCA). The reduced feature set is then used to train classifiers such as support vector machines (SVM), Gaussian Naive Bayes (GaussianNB), and random forests (RF) to differentiate between normal and fault conditions. Performance metrics, including precision, accuracy, recall, and F1-score, are computed for each combination of feature extractor and classifier. Extensive experiments with both the Solar Panel Images Dataset and the Infrared Solar Module Dataset show that the proposed approaches outperform state-of-the-art methods. For instance, the AIFD-SolDL approach utilizing SVM achieved perfect accuracy, precision, recall, and F1-score of 100% on the Solar Panel Images Dataset. Overall, the SPDA approach effectively detects faults, while deep learning techniques demonstrate high accuracy in fault classification, thereby enhancing the reliability of PV system maintenance and optimization.}, archive = {J_ASOC}, author = {Mohamed R. Shoaib and Heba M. Emara and Jun Zhao and Milad Taleby Ahvanooey and Essam Nabil}, doi = {10.1016/j.asoc.2025.113592}, journal = {Applied Soft Computing}, month = {12}, pages = {113592}, shortjournal = {Appl. Soft. Comput.}, title = {Advanced deep learning approaches for fault detection in solar PV systems: A comparative study of SPDA and AIFD-SolDL}, volume = {184}, year = {2025}, }
• (2025). Neuroevolution-based multiobjective algorithm for feature selection and binary classification of DNA microarrays. ASOC, 184, 113520. (www)

  @article{ , abstract = {The genomes of organisms have been sequenced for many years, leading to the discovery of thousands of genes. DNA microarrays are widely used tools for simultaneously analysing numerous genes, commonly employed in detecting and identifying various diseases, including cancer. However, microarray datasets have non-relevant and redundant information, hindering their analysis. This problem is further exacerbated considering the datasets’ high dimensionality and imbalanced classes. Consequently, standard practice involves incorporating a feature selection process to identify the most relevant genes and their associations with diseases. Various methods have been employed to address this task. However, none have taken a more holistic approach that effectively handles feature selection, automatically identifies the optimal classifier configuration, and manages potential conflicting objectives simultaneously. In response, this study introduces the S -metric selection - multiobjective neuroevolution of augmenting topologies (SMS-MONEAT) algorithm, which combines the multiobjective optimisation framework from S -metric selection - evolutionary multiobjective algorithm (SMS-EMOA) and the evolutionary operators from the neuroevolution algorithm N3O, a variation from NEAT which stands for ‘3 new operators’. SMS-MONEAT algorithm was designed to perform both feature selection and optimise the configuration of artificial neural networks for classification tasks. SMS-MONEAT was evaluated against classic and state-of-the-art methods for feature selection and microarray classification. The experiments were conducted on 20 highly challenging cancer-type datasets primarily sourced from the Curated Microarray Database, and the results were investigated for statistical significance. The findings suggest that SMS-MONEAT either outperforms or achieves competitive results in terms of classification compared to the mentioned methods, while at the same time, it selects a smaller subset of features.}, archive = {J_ASOC}, author = {Daniel García-Núñez and Katya Rodrígez-Vázquez and Carlos Hernández and Edgar Galván}, doi = {10.1016/j.asoc.2025.113520}, journal = {Applied Soft Computing}, month = {12}, pages = {113520}, shortjournal = {Appl. Soft. Comput.}, title = {Neuroevolution-based multiobjective algorithm for feature selection and binary classification of DNA microarrays}, volume = {184}, year = {2025}, }
• (2025). Deep learning-based bangla text normalization with emotion classification for expressive text-to-speech synthesizer. ASOC, 184, 112899. (www)

  @article{ , abstract = {This paper presents a novel method for text normalization and emotion classification specifically designed for the preprocessing steps of Expressive Text-to-Speech (ETTS) synthesizers. Text normalization, which converts non-standard words into standardized forms, is essential for generating clear and coherent output in ETTS synthesizers. Previous studies have encountered difficulties with Bangla homographic words, prompting the development of our proposed method. Our approach begins with the creation of a tokenized and categorized dataset consisting of 26 unique semiotic classes, utilizing a rule-based method with regular expressions. This dataset is derived from the Bangla Text Normalization Corpus (BTN Corpus), which contains 2 million Bangla sentences. Initially, Bangla written texts are tokenized, and each token is classified using a Temporal Convolutional Network (TCN) algorithm trained on the BTN Corpus to identify its semiotic class. This classification aids in generating normalized text, where the resulting tokens are reassembled into coherent sentences for final output. In addition to text normalization, we implement emotion classification for each normalized sentence using a Hierarchical Attention Network (HAN) model. The HAN model was trained on 67,277 normalized texts from the Bangla Normalized Emotion Text Corpus (BNET Corpus), categorizing each text into one of six emotion classes through a rule-based method with regular expressions. The proposed method demonstrates high accuracy rates, achieving a token classification accuracy of 99.977 % with the TCN and an emotion classification accuracy of 99.735 % with the HAN model.}, archive = {J_ASOC}, author = {Md. Rezaul Islam and Mohammad Shahidur Rahman}, doi = {10.1016/j.asoc.2025.112899}, journal = {Applied Soft Computing}, month = {12}, pages = {112899}, shortjournal = {Appl. Soft. Comput.}, title = {Deep learning-based bangla text normalization with emotion classification for expressive text-to-speech synthesizer}, volume = {184}, year = {2025}, }

AUTOM - 77

• (2026). Physics-informed learning-based distributed cooperative tracking control for multiple Euler–Lagrange systems. AUTOM, 183, 112653. (www)

  @article{ , abstract = {This article addresses the robust consensus tracking problem of multiple Euler–Lagrange systems subject to unknown dynamics and disturbances, in which only the root agent can access to the prescribed time-varying trajectory. A network structure that integrates the enhanced gate recurrent unit (GRU) and the physical-informed deep neural network (PI-DNN) is first developed to learn unknown dynamics. A distributed leader state estimator is designed and a distributed PI-DNN-GRU controller with weight adaptive law is further proposed. The closed-loop system is proved to achieve robust asymptotic consensus. Finally, numerical simulations demonstrate the effectiveness and advantages of the proposed method over the existing approaches.}, archive = {J_AUTOM}, author = {Kunfeng Yang and Huiping Li and Panfeng Huang}, doi = {10.1016/j.automatica.2025.112653}, journal = {Automatica}, month = {1}, pages = {112653}, shortjournal = {Automatica}, title = {Physics-informed learning-based distributed cooperative tracking control for multiple Euler–Lagrange systems}, volume = {183}, year = {2026}, }
• (2026). Reachable set estimation for cone-preserving linear systems with constrained exogenous inputs. AUTOM, 183, 112650. (www)

  @article{ , abstract = {This paper focuses on estimating the reachable set of cone-preserving linear systems subject to constrained exogenous inputs. Specifically, the input signal belongs to the space of essentially bounded vector-valued functions equipped with cone linear absolute norms, or to the space of vector-valued functions with the integral of their cone linear absolute norms being bounded. By adopting an auxiliary function composed of the maximum of a set of linear functions with respect to system states, sufficient conditions are provided to ensure that the reachable set is bounded by a given polyhedron. This treatment does not require the auxiliary function to be positive definite, and it is compatible with unstable systems, which shows a significant improvement over existing methods that use Lyapunov functions. Furthermore, we demonstrate that the proposed method is also suitable for delay systems. Finally, simulation examples are presented to verify the effectiveness of the proposed approach.}, archive = {J_AUTOM}, author = {Hongling Qiu and Jun Shen}, doi = {10.1016/j.automatica.2025.112650}, journal = {Automatica}, month = {1}, pages = {112650}, shortjournal = {Automatica}, title = {Reachable set estimation for cone-preserving linear systems with constrained exogenous inputs}, volume = {183}, year = {2026}, }
• (2026). Comment on “Anti-unwinding sliding mode attitude control via two modified rodrigues parameter sets for spacecraft” [Automatica 129 (2021) 109642]. AUTOM, 183, 112646. (www)

  @article{ , abstract = {This correspondence points out a flaw in Dong et al., (2021), making the proposed approach therein doubtful.}, archive = {J_AUTOM}, author = {Yuxin Su}, doi = {10.1016/j.automatica.2025.112646}, journal = {Automatica}, month = {1}, pages = {112646}, shortjournal = {Automatica}, title = {Comment on “Anti-unwinding sliding mode attitude control via two modified rodrigues parameter sets for spacecraft” [Automatica 129 (2021) 109642]}, volume = {183}, year = {2026}, }
• (2026). Unified PID control analysis for time-delay plants. AUTOM, 183, 112644. (www)

  @article{ , abstract = {In this paper, the gradient vector analysis and D-partition are combined to solve the general PID tuning problems for time-delay plants, including nominal stability, region poles placement, and stability margin constraint. This method can simultaneously tune two or three of four parameters: proportional, integral and differential gains as well as time-delay. Two key issues related to the determination of specified region in D-partition method can be solved using gradient analysis: 1. which direction of parameter variation on the boundary will reduce the number of unexpected closed-loop poles; 2. how many unexpected poles will change in the corresponding direction. By analyzing the geometric relationship between gradient vector and boundary, a unified method is developed to directly obtain the gradient vector at any boundary point without complicated calculation, and thus can find the parameter changing direction that can reduce the number of unexpected poles. Finally, the overall algorithm is proposed and two examples are provided to demonstrate the effectiveness of the proposed method. Our code ( Github ) is made publicly available for future researchers.}, archive = {J_AUTOM}, author = {Jiale An and Baitao Tang and Mingwei Sun and Jianhong Zhang and Zengqiang Chen}, doi = {10.1016/j.automatica.2025.112644}, journal = {Automatica}, month = {1}, pages = {112644}, shortjournal = {Automatica}, title = {Unified PID control analysis for time-delay plants}, volume = {183}, year = {2026}, }
• (2026). Distributed multi-time slot power balancing control of power systems with energy storage devices. AUTOM, 183, 112640. (www)

  @article{ , abstract = {This paper studies a crucial problem in power system balancing control, i.e., the multi-time slot economic dispatch (MTSED) problem, for power grids with substantial renewables, synchronous generators (SGs), and energy storage devices (ESDs). The target of MTSED is to optimally coordinate active/reactive power outputs of all controllable units to meet a forecast net demand profile over multiple time slots within a receding finite time horizon. Firstly, the MTSED is formulated as an optimization problem with operational constraints, including the limits on the output of each controllable unit, ramping rates of SGs, energy levels of ESDs, and bus voltages. Then, a novel projection-based algorithm is developed to solve the problem in a distributed way. In particular, the distributed algorithm is not limited to solving the MTSED problem but also applies to more general optimization problems with both generic convex objective functions and hard feasibility constraints. Finally, case studies verify the effectiveness of the proposed method.}, archive = {J_AUTOM}, author = {Luwei Yang and Xiemin Mo and Tao Liu and David J. Hill}, doi = {10.1016/j.automatica.2025.112640}, journal = {Automatica}, month = {1}, pages = {112640}, shortjournal = {Automatica}, title = {Distributed multi-time slot power balancing control of power systems with energy storage devices}, volume = {183}, year = {2026}, }
• (2026). A conditional invertible neural network-based particle filter. AUTOM, 183, 112639. (www)

  @article{ , abstract = {This study proposes a neural network-based approach to improve the efficiency of particle filters. A conditional invertible neural network (CINN) is employed to approximate the optimal proposal distribution in a data-driven manner. The CINN is trained offline to enhance the accuracy of conditional state inference, thereby enabling the generation of high-quality particles and improving the overall filtering performance. During online filtering, the CINN serves as the proposal mechanism for particle updates. The architecture is designed based on two key principles: (1) enabling efficient particle propagation and (2) allowing for closed-form computation of particle weights. The effectiveness of the proposed particle filter framework is validated through three numerical examples.}, archive = {J_AUTOM}, author = {Wenxin Sun and Weili Xiong and Biao Huang and Hongtian Chen}, doi = {10.1016/j.automatica.2025.112639}, journal = {Automatica}, month = {1}, pages = {112639}, shortjournal = {Automatica}, title = {A conditional invertible neural network-based particle filter}, volume = {183}, year = {2026}, }
• (2026). Angle-based formation stabilization and maneuvers in port-hamiltonian form with bearing and velocity measurements. AUTOM, 183, 112638. (www)

  @article{ , abstract = {This paper proposes a port-Hamiltonian approach for angle-based formation stabilization and maneuvers using bearing and velocity measurements with an underlying triangulated Laman graph. The corresponding port-Hamiltonian controller is designed using virtual couplings on the errors of angle constraints in angle space. Then, the angle constraints and agent actuators are mapped by the constraint Jacobian, which can be applied to other formation constraints. In addition, to avoid unavailable distance terms in the control law, an estimator is designed based on port-Hamiltonian theory and the property that energy is coordinate-free for different sensor modalities using bearing and velocity measurements, which permits our framework to inject damping for the formation maneuvers. Furthermore, the stability analysis for several maneuvers is provided. Simulations are performed to illustrate the effectiveness of the approach.}, archive = {J_AUTOM}, author = {Ningbo Li and Pablo Borja and Arjan van der Schaft and Jacquelien M.A. Scherpen}, doi = {10.1016/j.automatica.2025.112638}, journal = {Automatica}, month = {1}, pages = {112638}, shortjournal = {Automatica}, title = {Angle-based formation stabilization and maneuvers in port-hamiltonian form with bearing and velocity measurements}, volume = {183}, year = {2026}, }
• (2026). An alternative paradigm of fault diagnosis in dynamic systems: Orthogonal projection-based methods. AUTOM, 183, 112637. (www)

  @article{ , abstract = {In this paper, a paradigm of fault diagnosis in dynamic systems is proposed as an alternative to the well-established observer-based framework. The basic idea behind this work is to deal with fault detection and isolation by projecting data onto system subspaces in Hilbert space with the aid of orthogonal projection operators and using gap metric for threshold setting. According to the Hilbert Projection Theorem, an orthogonal projection leads to a unique decomposition of the data into nominal and uncertain ones and enables the use of gap metric as the similarity measure. As a result, the projection-based fault diagnosis systems are capable to handle uncertainties and multiplicative faults efficiently, facilitate uniformly addressing residual generation and threshold setting, and deliver optimal fault detectability. As a part of our work, relations to the existing observer-based fault detection systems are investigated. The experimental study on a three-tank system showcases the capability of the proposed projection-based fault diagnosis methods.}, archive = {J_AUTOM}, author = {Steven X. Ding and Linlin Li and Tianyu Liu}, doi = {10.1016/j.automatica.2025.112637}, journal = {Automatica}, month = {1}, pages = {112637}, shortjournal = {Automatica}, title = {An alternative paradigm of fault diagnosis in dynamic systems: Orthogonal projection-based methods}, volume = {183}, year = {2026}, }
• (2026). Congealed neural network design for uncertain nonlinear spatiotemporal control systems. AUTOM, 183, 112636. (www)

  @article{ , abstract = {This note is concerned with the approximation-based adaptive control problem for a class of uncertain nonlinear spatiotemporal systems. A congealed neural network (ConNN) is first proposed to approximate nonlinear spatiotemporal uncertainties arising from system states and time-varying parameters. Unlike conventional NN approximation structures, the ConNN explicitly decomposes the time-varying coupling weight into a congealed weight and a time-dependent perturbation. The congealed weight is estimated using a standard adaptive law for constant parameters, while the residual perturbation is handled within the network structure. To enhance robustness, smooth sliding-mode-like functions are then embedded into the control architecture, effectively attenuating bias terms, particularly in reference tracking scenarios. It is shown that the resulting ConNN-based adaptive controller guarantees adjustable, bounded-error tracking performance, thereby extending the applicability of robust adaptive control to complex spatiotemporal systems and outperforming existing NN-based approaches.}, archive = {J_AUTOM}, author = {Tianrun Liu and Yang-Yang Chen and Xiaohua Ge}, doi = {10.1016/j.automatica.2025.112636}, journal = {Automatica}, month = {1}, pages = {112636}, shortjournal = {Automatica}, title = {Congealed neural network design for uncertain nonlinear spatiotemporal control systems}, volume = {183}, year = {2026}, }
• (2026). Stabilization of feedforward nonlinear time-delay systems with vanishing actuator effectiveness by linear time-varying feedback. AUTOM, 183, 112635. (www)

  @article{ , abstract = {This paper considers the problem of global stabilization for a class of feedforward nonlinear systems in the presence of state delay, uncertain parameters, and vanishing actuator effectiveness. By assuming that the time-delay nonlinear function is unknown but satisfies a linear growth condition with unknown parameters, a state feedback controller incorporating two time-varying parameters is designed. One parameter focuses on compensating for the vanishing actuator effectiveness, while the other addresses uncertain nonlinear terms. The feature of the approach is that the designed controller is linear, and its implementation do not necessitate online computation, thereby significantly conserving computational resources. An example is included to verify the effectiveness of the proposed approaches.}, archive = {J_AUTOM}, author = {Kai Zhang and Bin Zhou}, doi = {10.1016/j.automatica.2025.112635}, journal = {Automatica}, month = {1}, pages = {112635}, shortjournal = {Automatica}, title = {Stabilization of feedforward nonlinear time-delay systems with vanishing actuator effectiveness by linear time-varying feedback}, volume = {183}, year = {2026}, }
• (2026). Distributed safety-critical control of nonlinear multi-agent systems. AUTOM, 183, 112634. (www)

  @article{ , abstract = {This paper considers the safety-critical control problem for nonlinear second-order multi-agent systems with constraints of each agent and inter-agent ones. We overcome the challenge of the time-varying and position-dependent communication network with limited sensing range by introducing a truncated function for the smooth addition and deletion of links in the edge set, and design a distributed and locally Lipschitz-continuous safety-critical control law, composed of a nominal controller for the objectives such as consensus, formation, and position swapping, etc., and a safety controller, which only takes effect when some neighboring agent enters the custom-designed boundary set. Meanwhile, to rigorously verify the safety of the whole multi-agent system, a continuously differentiable control barrier function is proposed under a relaxed feasibility condition in the sense that it is imposed on each subsystem and only needed in the boundary area.}, archive = {J_AUTOM}, author = {Xiaoyu Wang and Yi Dong and Yiguang Hong and Karl Henrik Johansson}, doi = {10.1016/j.automatica.2025.112634}, journal = {Automatica}, month = {1}, pages = {112634}, shortjournal = {Automatica}, title = {Distributed safety-critical control of nonlinear multi-agent systems}, volume = {183}, year = {2026}, }
• (2026). Integral version of Lyapunov–Razumikhin conditions. AUTOM, 183, 112633. (www)

  @article{ , abstract = {For a retarded nonlinear system, this paper proposes a modification of the Lyapunov–Razumikhin method introducing a distributed integral of Lyapunov function for comparison with its current value instead of the maximum over delayed interval. The developed conditions are illustrated on different analysis problems presenting integral Halanay-type inequalities.}, archive = {J_AUTOM}, author = {Jian Wang and Denis Efimov and Alexander Aleksandrov}, doi = {10.1016/j.automatica.2025.112633}, journal = {Automatica}, month = {1}, pages = {112633}, shortjournal = {Automatica}, title = {Integral version of Lyapunov–Razumikhin conditions}, volume = {183}, year = {2026}, }
• (2026). Distributed adaptive fixed-time formation tracking for heterogeneous multi-agent systems. AUTOM, 183, 112632. (www)

  @article{ , abstract = {This article investigates the problem of fixed-time time-varying formation tracking control (FTC) for heterogeneous linear multiagent systems (MASs) under the directed communication graph. It is assumed that the Laplacian matrix associated with the communication graph is unavailable and that the system matrices of the leader are only available to its neighboring followers. This differs from many existing works on fixed-time FTC problems where the communication graphs are typically undirected and protocol designs often rely on certain global information. A novel distributed observer is first put forward to estimate both the state and system matrices of the leader in fixed time. Then, an adaptive scheme is developed to solve the time-varying regulator equations resulting from the estimated leader system matrices in fixed time. Based on the proposed observer and the adaptive solutions to the regulator equations, a distributed adaptive fixed-time FTC protocol is further proposed via coordinate transformation techniques. It is shown that our proposed controllers do not require the input matrices of the followers to be of full row rank. It is also shown that the concerned fixed-time FTC problem can be solved with the proposed fixed-time FTC strategy in a distributed manner. Our results can be directly applied to solve both the adaptive fixed-time cooperative output regulation problem and the leader-following consensus problems of MASs under the directed graph. Finally, the effectiveness of the proposed fixed-time FTC strategy is demonstrated through a numerical example.}, archive = {J_AUTOM}, author = {Shiyu Zhou and Dong Sun and Gang Feng}, doi = {10.1016/j.automatica.2025.112632}, journal = {Automatica}, month = {1}, pages = {112632}, shortjournal = {Automatica}, title = {Distributed adaptive fixed-time formation tracking for heterogeneous multi-agent systems}, volume = {183}, year = {2026}, }
• (2026). Efficient iterative deadlock prevention for flexible manufacturing systems utilizing modified resource requirement graphs. AUTOM, 183, 112631. (www)

  @article{ , abstract = {This article deals with deadlock problems in flexible manufacturing systems (FMSs) modeled by Petri nets. An efficient iterative technique is proposed to synthesize a liveness-enforcing supervisor composed of control places. The main idea behind the proposed approach is to construct at each iteration a digraph representation of a Petri net called a modified resource requirement graph (MR 2 G), which serves as an enhanced version of a resource requirement graph (RRG). The number and type of resources required by a process can be directly obtained from the MR 2 G. Subsequently, an integer linear programming problem (ILPP) is formulated by exploiting the MR 2 G in order to identify a partial deadlock (if any) using an objective function that minimizes the sum of tokens in operation places at the obtained partial deadlock. A place invariant is then designed to compute a control place that prevents the identified partial deadlock from being reached. Due to the insertion of a new control place, it is necessary to augment the MR 2 G accordingly, and iterate (by adding new inequalities to the ILPP and adjusting the existing ones). This iterative process is carried out until no partial deadlock can be identified by the ILPP. The developed approach is computationally competitive and applicable to large-scale systems since both the number of constraints and the number of variables in the ILPP grow only polynomially with the size of the given Petri net model. Several FMS examples illustrate the applicability and effectiveness of the proposed method.}, archive = {J_AUTOM}, author = {Yao Lu and Yufeng Chen and Christoforos N. Hadjicostis and Zhiwu Li}, doi = {10.1016/j.automatica.2025.112631}, journal = {Automatica}, month = {1}, pages = {112631}, shortjournal = {Automatica}, title = {Efficient iterative deadlock prevention for flexible manufacturing systems utilizing modified resource requirement graphs}, volume = {183}, year = {2026}, }
• (2026). Estimation of the minimum and maximum states of charge of lithium-ion battery packs: A hybrid approach. AUTOM, 183, 112630. (www)

  @article{ , abstract = {Monitoring the minimum and maximum states of charge (SOC) in lithium-ion battery packs is key to ensuring safe and reliable long-term operation. The challenge is that these SOCs cannot be directly measured and their corresponding cells within the pack may change with time. This paper proposes a novel hybrid scheme that estimates the minimum and maximum SOCs within a battery pack given by the series interconnection of equivalent circuit models. The estimation scheme relies on a mechanism that determines online two cells, which are candidates for having the minimum and maximum SOCs. The dimension of the hybrid estimator is independent of the number of cells, which makes it particularly attractive for large battery packs. Moreover, the estimator is endowed with robust, global convergence guarantees despite disturbances and measurement noise. Furthermore, Zeno behavior is ruled out as any solution to the considered system is shown to exhibit an average dwell-time. Numerical simulations illustrate the efficiency in terms of accuracy and computational time of the proposed estimator.}, archive = {J_AUTOM}, author = {Mira Khalil and Romain Postoyan and Stéphane Raël and Dragan Nešić}, doi = {10.1016/j.automatica.2025.112630}, journal = {Automatica}, month = {1}, pages = {112630}, shortjournal = {Automatica}, title = {Estimation of the minimum and maximum states of charge of lithium-ion battery packs: A hybrid approach}, volume = {183}, year = {2026}, }
• (2026). Nature-inspired dynamic control for pursuit-evasion of robots. AUTOM, 183, 112629. (www)

  @article{ , abstract = {The pursuit-evasion problem is widespread in nature, engineering, and societal applications. It is commonly observed in nature that predators often exhibit faster speeds than their prey but have less agile maneuverability. Over millions of years of evolution, animals have developed effective and efficient strategies for pursuit and evasion. In this paper, we provide a dynamic framework for the pursuit-evasion problem of unicycle systems, drawing inspiration from nature. First, we address the scenario with one pursuer and one evader by proposing an Alert-Turn control strategy, which consists of two efficient ingredients: a sudden turning maneuver and an alert condition for starting and maintaining the maneuver. We present and analyze the escape and capture results at two levels: a lower level of a single run and a higher level with respect to parameters’ changes. In addition, we provide a theorem with sufficient conditions for capture. The Alert-Turn strategy is then extended to more complex scenarios involving multiple pursuers and evaders by integrating aggregation control laws and a target-changing mechanism. By adjusting a ‘selfish parameter’, the aggregation control commands produce various escape patterns of evaders: cooperative mode, selfish mode, and their combinations. The influence of the selfish parameter is quantified, and the target-changing mechanism is explored from a statistical perspective. Our findings align closely with observations in nature. Finally, the proposed strategies are validated through numerical simulations that replicate some chasing behaviors of animals in nature.}, archive = {J_AUTOM}, author = {Panpan Zhou and Sirui Li and Benyun Zhao and Bo Wahlberg and Xiaoming Hu}, doi = {10.1016/j.automatica.2025.112629}, journal = {Automatica}, month = {1}, pages = {112629}, shortjournal = {Automatica}, title = {Nature-inspired dynamic control for pursuit-evasion of robots}, volume = {183}, year = {2026}, }
• (2026). Event/self-triggered cooperative control via reinforcement learning for a quadrotor team under multiple faults and denial-of-service attacks. AUTOM, 183, 112628. (www)

  @article{ , abstract = {The cooperative fault-tolerant control problem of a quadrotor team, subject to multiple faults and denial-of-service (DoS) attacks, is addressed via event/self-triggered strategies and reinforcement learning. Multiple under-actuated quadrotors with nonlinearities, couplings, and unknown dynamical parameters are introduced to achieve distributed cooperation. Event-triggered observers are developed to estimate position references under uncertain cyber faults and DoS attacks. Observer-based optimal controllers for the nominal position and attitude subsystems are learned by off-policy reinforcement learning without dynamical information. Fault-tolerant controllers are constructed by integrating the learned controllers and adaptive actuator fault compensators. A self-triggered strategy is further proposed to avoid continuous communication. The relationship among the time-varying topology coupled with cyber faults, the features of DoS attacks, and the closed-loop control system stability is analyzed, and the exclusion of Zeno behavior is proved. Simulation results illustrate the effectiveness of the proposed methods.}, archive = {J_AUTOM}, author = {Ziming Ren and Hao Liu and Hongwei Zhang and Ci Chen and Frank L. Lewis}, doi = {10.1016/j.automatica.2025.112628}, journal = {Automatica}, month = {1}, pages = {112628}, shortjournal = {Automatica}, title = {Event/self-triggered cooperative control via reinforcement learning for a quadrotor team under multiple faults and denial-of-service attacks}, volume = {183}, year = {2026}, }
• (2026). Convergence behaviors of nonlinear polar opinion dynamics in cooperative-antagonistic social networks. AUTOM, 183, 112627. (www)

  @article{ , abstract = {This paper investigates two kinds of convergence behaviors of nonlinear polar opinion dynamics over cooperative-antagonistic social networks determined by signed graphs. The first problem studies the neutralization of opinions over strongly connected and structurally unbalanced signed graphs, while the second problem focuses on the containment convergence to external stubborn agents over signed graphs under a variety of connectivity constraints. It is observed that various connectivity conditions induce the common intrinsic property that the involved generalized opposing Laplacian matrix is nonsingular. A stability framework is proposed that addresses both the existence of potential limiting behaviors as well as their convergence properties. This involves an analysis of the positiveness of the (opinion) susceptibility functions. The domains of attraction associated with the convergent behaviors are characterized. The solvability of both problems is provided, and the attracting domains are characterized precisely. The results are applied to blended social networks composed of three specialized opinion models. Simulations and comparisons are provided to illustrate the effectiveness of the paper’s results.}, archive = {J_AUTOM}, author = {Ti-Chung Lee and Ying Zhang and Youfeng Su and Iven Mareels}, doi = {10.1016/j.automatica.2025.112627}, journal = {Automatica}, month = {1}, pages = {112627}, shortjournal = {Automatica}, title = {Convergence behaviors of nonlinear polar opinion dynamics in cooperative-antagonistic social networks}, volume = {183}, year = {2026}, }
• (2026). Linear system analysis and optimal control of natural gas dynamics in pipeline networks. AUTOM, 183, 112626. (www)

  @article{ , abstract = {We design nonlinear and adaptive linear model-predictive control (MPC) techniques to minimize operational costs of compressor-actuated dynamics in natural gas pipeline networks. We establish stability of the local linear system and derive rigorous bounds on error between the nonlinear and linear system solutions. These bounds are used to quantify conditions under which the linear MPC can substitute the nonlinear MPC without significant loss of predictive accuracy. Furthermore, we prove and numerically verify that the computational cost of the linear MPC is orders of magnitude lower than that of solving the baseline optimal control problem. Numerical simulations are performed on nontrivial networks to demonstrate that the proposed MPC can effectively adapt to varying load conditions while maintaining nearly 95% optimality.}, archive = {J_AUTOM}, author = {Luke S. Baker and Sachin Shivakumar and Dieter Armbruster and Rodrigo B. Platte and Anatoly Zlotnik}, doi = {10.1016/j.automatica.2025.112626}, journal = {Automatica}, month = {1}, pages = {112626}, shortjournal = {Automatica}, title = {Linear system analysis and optimal control of natural gas dynamics in pipeline networks}, volume = {183}, year = {2026}, }
• (2026). Adaptive distributed observer design for nonlinear multiagent systems. AUTOM, 183, 112625. (www)

  @article{ , abstract = {Distributed state estimation is crucial for the leader-following control problem of multiagent systems (MASs). In this paper, adaptive distributed observers (DOs) are designed for a nonlinear autonomous leader system using only its output. Via system transformation, the DO design problem of the origin system is converted to that of a canonical system with lumped dynamics. When the lumped dynamics is parametric, an adaptive DO is developed to reconstruct the state and the unknown parameters under an undirected topology, which also addresses the distributed state/parameter estimation problem of an uncertain autonomous system with its dynamics in a parametric form. Then, the DO framework is extended to the case of non-parametric uncertainties, and a neural network (NN) DO is designed for reconstructing the state/lumped dynamics over a strongly connected digraph Finally, the effectiveness of the proposed DOs is demonstrated via numerical simulations.}, archive = {J_AUTOM}, author = {Jixing Lv and Changhong Wang and Lihua Xie}, doi = {10.1016/j.automatica.2025.112625}, journal = {Automatica}, month = {1}, pages = {112625}, shortjournal = {Automatica}, title = {Adaptive distributed observer design for nonlinear multiagent systems}, volume = {183}, year = {2026}, }
• (2026). Simple sufficient conditions for integer stabilizability of discrete-time systems with relative degree one. AUTOM, 183, 112624. (www)

  @article{ , abstract = {In this paper, firstly the concept of integer stabilizability is defined. It is discussed that the main motivation for introducing this concept is its helpfulness in achieving homomorphic encrypted control systems involving no computational overflow. Then, some simple sufficient conditions to ensure integer stabilizability of discrete-time systems described by transfer functions with relative degree one are derived. The usefulness of achievements of the paper is verified by a numerical example.}, archive = {J_AUTOM}, author = {Mohammad Saleh Tavazoei}, doi = {10.1016/j.automatica.2025.112624}, journal = {Automatica}, month = {1}, pages = {112624}, shortjournal = {Automatica}, title = {Simple sufficient conditions for integer stabilizability of discrete-time systems with relative degree one}, volume = {183}, year = {2026}, }
• (2026). A novel switching rule to observer-based control for switched systems. AUTOM, 183, 112622. (www)

  @article{ , abstract = {This study explores observer-based control for switched systems subject to partially unknown sojourn probability information. To more accurately capture the dynamic nature of switched systems, a novel switching rule related to duration time is devised, in which the system switching is governed by a joint distribution function. This function uniquely integrates both the current system mode and its duration, effectively addressing the challenges of incompleteness in probability information. Distinct from the conventional semi-Markov kernel technique, the sojourn probabilities involve fewer parameters and can be more readily measured using statistical techniques, thereby lessening computational burden. Additionally, a non-monotonic Lyapunov function is constructed, leading to less conservative conditions that ensure the mean-square stability of the switched systems. Eventually, the effectiveness and superiority of the devised methodology are verified through a practical example.}, archive = {J_AUTOM}, author = {Jun Cheng and Tianfeng Tang and Huaicheng Yan and Zheng-Guang Wu and Dan Zhang}, doi = {10.1016/j.automatica.2025.112622}, journal = {Automatica}, month = {1}, pages = {112622}, shortjournal = {Automatica}, title = {A novel switching rule to observer-based control for switched systems}, volume = {183}, year = {2026}, }
• (2026). Sampled-data primal–dual gradient dynamics in model predictive control. AUTOM, 183, 112621. (www)

  @article{ , abstract = {Model predictive control (MPC) can incur computational burden that exceeds what the application warrants, sometimes even for linear systems. Recently, a rapid computation method that guides the input toward convergence with the optimal control problem solution by employing primal–dual gradient (PDG) dynamics has been proposed for linear MPCs. However, stability has been ensured under the assumption that the controller is a continuous-time system, leading to potential instability when the controller undergoes discretization and is implemented as a sampled-data system. In this paper, we propose a discrete-time dynamical controller incorporating specific modifications to the PDG approach and present stability conditions. Additionally, we introduce an extension to enhance control performance that was traded off in the original. Numerical examples substantiate that our proposed method, which can be executed in only 1 μ s on a commodity laptop, not only ensures stability considering sampled-data implementation but also substantially enhances the control performance.}, archive = {J_AUTOM}, author = {Ryuta Moriyasu and Sho Kawaguchi and Kenji Kashima}, doi = {10.1016/j.automatica.2025.112621}, journal = {Automatica}, month = {1}, pages = {112621}, shortjournal = {Automatica}, title = {Sampled-data primal–dual gradient dynamics in model predictive control}, volume = {183}, year = {2026}, }
• (2026). Cooperative dynamic surrounding coverage control for multi-autonomous surface vehicle fleets with limited sensing ranges. AUTOM, 183, 112620. (www)

  @article{ , abstract = {Cooperative surrounding coverage by a multi-autonomous surface vehicle (ASV) fleet involves dynamic coverage with concentric moving trajectories. This low-cost form of coverage, which operates with limited perceptual ranges, has long presented a significant challenge in real-world collective marine detection applications. To this end, the paper establishes a surrounding coverage control framework that minimizes the required number of ASVs by covering a designated region across multiple consecutive periods. A distributed randomized gradient-free (DRGF) protocol is thereby proposed to identify the optimal central point of the coverage region. Accordingly, a distributed control law is designed to regulate the ASV dynamic surrounding coverage operation around the optimal concentric point. Significantly, sufficient conditions are derived to guarantee the asymptotical convergence of the closed-loop multi-ASV fleet governed by the present DGRF. Finally, experiments are conducted on a self-established lake-based multi-ASV fleet platform to verify the effectiveness of the present surrounding coverage scheme.}, archive = {J_AUTOM}, author = {Jiayu Zou and Hai-Tao Zhang and Bin Liu and Xiaohua Liu and Jianing Ding and Ning Xing and Shiqi Fu and Lijun Zhu}, doi = {10.1016/j.automatica.2025.112620}, journal = {Automatica}, month = {1}, pages = {112620}, shortjournal = {Automatica}, title = {Cooperative dynamic surrounding coverage control for multi-autonomous surface vehicle fleets with limited sensing ranges}, volume = {183}, year = {2026}, }
• (2026). Asynchronous resilient collision-free formation control for nonlinear MASs under DoS attacks. AUTOM, 183, 112619. (www)

  @article{ , abstract = {The multi-agent systems (MASs) exchange the communication information by the communication network, which is vulnerable to denial-of-service (DoS) attacks. DoS attacks can disrupt the communication channels and result in the collisions between agents, thus it is an important and challenging topic to study the collision avoidance formation control issue under DoS attacks. This paper investigates the asynchronous resilient collision-free formation control issue for nonlinear MASs with DoS attacks. An asynchronous resilient distributed observer is proposed to estimate the output information of leader under DoS attacks. Based on the designed distributed observer and by constructing two high-order filters, an asynchronous resilient collision-free formation control method is presented by backstepping control design theory. The presented formation control method can ensure that the formation tracking errors asymptotically converge to zero, and the collision avoidance objective is realized. Moreover, the non-differentiable problem of virtual controllers is solved. Finally, we apply the developed formation control method to multiple Euler-Lagrangian (EL) systems, the simulation and comparison results verify its effectiveness.}, archive = {J_AUTOM}, author = {Jun Zhang and Jun Ning and Shaocheng Tong}, doi = {10.1016/j.automatica.2025.112619}, journal = {Automatica}, month = {1}, pages = {112619}, shortjournal = {Automatica}, title = {Asynchronous resilient collision-free formation control for nonlinear MASs under DoS attacks}, volume = {183}, year = {2026}, }
• (2026). Sector stabilization criterion of a novel nonlinear flexible marine riser coupled system. AUTOM, 183, 112618. (www)

  @article{ , abstract = {This paper establishes a sector stabilization criterion for a nonlinear flexible marine riser system that incorporates lateral and transverse coupling vibrations, derived from Hamilton’s principle. This criterion, grounded in the sector-bounded condition, encompasses a wide range of linear and nonlinear feedback control laws applied to the transverse and lateral directions at the top boundary of the flexible marine riser, respectively. In the analysis, the nonlinear semigroup theory is utilized to establish the well-posedness of the resulting closed-loop coupled system. Notably, the solution demonstrates continuous dependence on the initial conditions. Furthermore, the exponential stability of the closed-loop coupled system is achieved by employing a generalized Gronwall-type integral inequality and the integral multiplier method, which involves the innovative development of an energy-like functional. To demonstrate the effectiveness of the proposed controls, numerical simulations utilizing the finite element method are presented.}, archive = {J_AUTOM}, author = {Yi Cheng and Xin Wang and Yuhu Wu and Bao-Zhu Guo}, doi = {10.1016/j.automatica.2025.112618}, journal = {Automatica}, month = {1}, pages = {112618}, shortjournal = {Automatica}, title = {Sector stabilization criterion of a novel nonlinear flexible marine riser coupled system}, volume = {183}, year = {2026}, }
• (2026). Global exponential stabilization of 2 × 2 linear hyperbolic PDEs via dynamic event-triggered backstepping control. AUTOM, 183, 112617. (www)

  @article{ , abstract = {This paper introduces novel dynamic event-triggered control (ETC) mechanisms for 2 × 2 linear hyperbolic PDEs in three configurations: continuous-time event-triggered control (CETC), periodic event-triggered control (PETC), and self-triggered control (STC). These mechanisms ensure global exponential stability (GES) under ETC using PDE backstepping, with stability estimates provided in the spatial L 2 norm of the states. The proposed CETC and PETC designs are observer-based and require continuous boundary measurements collocated with the actuation. In contrast, the STC design requires full-state measurements; however, unlike CETC and PETC, it does not require continuous measurements for the triggering mechanism—only measurements taken at event times. In the CETC design, a lower bound on the time between two consecutive events is enforced, and a dynamic variable with appropriately designed switching dynamics is introduced. By employing a novel Lyapunov functional, GES of the closed-loop system is established under zero-order hold implementation of the backstepping control between events. Events are triggered when the dynamic variable crosses zero from the positive side, after which it is immediately reset to an appropriate nonnegative value. Detecting events, therefore, necessitates continuous monitoring of this dynamic variable. To address this limitation, PETC and STC strategies are proposed. The PETC design identifies a suitable triggering condition that requires only periodic checks and derives an upper bound on the allowable sampling period. This PETC approach preserves the GES guaranteed by CETC without requiring continuous monitoring of a triggering condition, although it still relies on continuous measurements. Unlike CETC and PETC, STC requires neither continuous measurements nor monitoring of a triggering condition. Instead, at each event, STC computes the time to the next event — beyond a suitably enforced minimal dwell-time — using only measurements taken at events. Despite relying solely on event-triggered measurements, STC is capable of guaranteeing GES of the closed-loop system. The well-posedness of the closed-loop systems under all three strategies is established. A simulation study is provided to illustrate the theoretical results.}, archive = {J_AUTOM}, author = {Bhathiya Rathnayake and Mamadou Diagne}, doi = {10.1016/j.automatica.2025.112617}, journal = {Automatica}, month = {1}, pages = {112617}, shortjournal = {Automatica}, title = {Global exponential stabilization of 2 × 2 linear hyperbolic PDEs via dynamic event-triggered backstepping control}, volume = {183}, year = {2026}, }
• (2026). Signal temporal logic control synthesis among uncontrollable dynamic agents with conformal prediction. AUTOM, 183, 112616. (www)

  @article{ , abstract = {The control of dynamical systems under temporal logic specifications among uncontrollable dynamic agents is challenging due to the agents’ a-priori unknown behavior. Existing works have considered the problem where either all agents are controllable, the agent models are deterministic and known, or no safety guarantees are provided. We propose a predictive control synthesis framework that guarantees, with high probability, the satisfaction of signal temporal logic (STL) tasks that are defined over a controllable system in the presence of uncontrollable stochastic agents. We use trajectory predictors and conformal prediction to construct probabilistic prediction regions for each uncontrollable agent that are valid over multiple future time steps. Specifically, we construct a normalized prediction region over all agents and time steps to reduce conservatism and increase data efficiency. We then formulate a worst-case bilevel mixed integer program (MIP) that accounts for all agent realizations within the prediction region to obtain an open-loop controller that provably guarantee task satisfaction with high probability. To efficiently solve this bilevel MIP, we propose an equivalent MIP program based on KKT conditions of the original bilevel formulation. Building upon this, we design a closed-loop controller, where both recursive feasibility and task satisfaction can be guaranteed with high probability. We illustrate our control synthesis framework on two case studies.}, archive = {J_AUTOM}, author = {Xinyi Yu and Yiqi Zhao and Xiang Yin and Lars Lindemann}, doi = {10.1016/j.automatica.2025.112616}, journal = {Automatica}, month = {1}, pages = {112616}, shortjournal = {Automatica}, title = {Signal temporal logic control synthesis among uncontrollable dynamic agents with conformal prediction}, volume = {183}, year = {2026}, }
• (2026). Event-triggered boundary estimation for 2 × 2 hyperbolic PDEs with disturbance. AUTOM, 183, 112615. (www)

  @article{ , abstract = {This paper addresses the event-triggered boundary estimation problem for 2 × 2 linear hyperbolic partial differential equations (PDEs) subject to disturbances. These disturbances enter the PDEs through the boundary and are represented as an ordinary differential equation (ODE). The scenario considered involves transmitting boundary measurements to the observer only when necessary, as determined by a dynamic event-triggering condition. This approach aims to conserve communication and computational resources. The event-triggered observer is designed using the backstepping method, and an event-triggering condition is proposed to determine the time instants when sampled measurements should be transmitted. Under this event-triggered estimation mechanism, a minimal dwell-time between consecutive triggering time instants is guaranteed. Furthermore, the well-posedness of the solutions and the exponential convergence of the estimation error are ensured. Finally, the obtained results are applied to the Saint-Venant equations to demonstrate the effectiveness of the event-triggered estimation method.}, archive = {J_AUTOM}, author = {Yan Zhao and Rafael Vazquez}, doi = {10.1016/j.automatica.2025.112615}, journal = {Automatica}, month = {1}, pages = {112615}, shortjournal = {Automatica}, title = {Event-triggered boundary estimation for 2 × 2 hyperbolic PDEs with disturbance}, volume = {183}, year = {2026}, }
• (2026). Distributed data-driven unknown-input observers. AUTOM, 183, 112614. (www)

  @article{ , abstract = {Unknown inputs related to, e.g., sensor aging, modeling errors, or device bias, represent a major concern in wireless sensor networks, as they degrade the state estimation performance. To improve the performance, unknown-input observers (UIOs) have been proposed. Most of the results available to design UIOs are based on explicit system models, which can be difficult or impossible to obtain in real-world applications. Data-driven techniques, on the other hand, have become a viable alternative for the design and analysis of unknown systems using only data. In this context, a novel data-driven distributed unknown-input observer (D-DUIO) for unknown continuous-time linear time-invariant (LTI) systems is developed, which requires solely some data collected offline, without any prior knowledge of the system matrices. In the paper, first, a model-based approach to the design of a DUIO is presented. A sufficient condition for the existence of such a DUIO is recalled, and a new one is proposed, that is prone to a data-driven adaptation. Moving to a data-driven approach, it is shown that under suitable assumptions on the input/output/state data collected from the continuous-time system, it is possible to both claim the existence of a D-DUIO and to derive its matrices in terms of the matrices of pre-collected data. Finally, the efficacy of the D-DUIO is illustrated by means of numerical examples.}, archive = {J_AUTOM}, author = {Yuzhou Wei and Giorgia Disarò and Wenjie Liu and Jian Sun and Maria Elena Valcher and Gang Wang}, doi = {10.1016/j.automatica.2025.112614}, journal = {Automatica}, month = {1}, pages = {112614}, shortjournal = {Automatica}, title = {Distributed data-driven unknown-input observers}, volume = {183}, year = {2026}, }
• (2026). Robustness of supervisory controllers subject to measurement disturbances. AUTOM, 183, 112613. (www)

  @article{ , abstract = {This paper studies the supervisory control problem for a class of uncertain nonlinearly parameterized systems in the presence of measurement disturbance. Based on the well-established estimator-based supervisory control structure, a robustification of supervisory controller dealing with measurement disturbance is developed taking advantage of the monitoring signals redesign, such that the plant state asymptotically converges to a given set-point or its neighborhood, subject to the vanishing or the persistent measurement disturbance, respectively. Moreover, a constructive design of the multi-estimator using measurement feedback is proposed for the supervisory control of a class of uncertain nonlinearly parameterized systems in the strict-feedback form. A numerical simulation based on an uncertain mass–spring system is given to show the efficacy of our proposed algorithm, in which we use an event-trigger to be a measurement disturbance generator.}, archive = {J_AUTOM}, author = {Yutian Wang and Qingkai Meng and Yi Jiang}, doi = {10.1016/j.automatica.2025.112613}, journal = {Automatica}, month = {1}, pages = {112613}, shortjournal = {Automatica}, title = {Robustness of supervisory controllers subject to measurement disturbances}, volume = {183}, year = {2026}, }
• (2026). Data-driven min–max MPC for linear systems: Robustness and adaptation. AUTOM, 183, 112612. (www)

  @article{ , abstract = {Data-driven controllers design is an important research problem, in particular when data is corrupted by the noise. In this paper, we propose a data-driven min–max model predictive control (MPC) scheme using noisy input-state data for unknown linear time-invariant (LTI) system. The unknown system matrices are characterized by a set-membership representation using the noisy input-state data. Leveraging this representation, we derive an upper bound on the worst-case cost and determine the corresponding optimal state-feedback control law through a semidefinite program (SDP). We prove that the resulting closed-loop system is robustly stabilized and satisfies the input and state constraints. Further, we propose an adaptive data-driven min–max MPC scheme which exploits additional online input-state data to improve closed-loop performance. Numerical examples show the effectiveness of the proposed methods.}, archive = {J_AUTOM}, author = {Yifan Xie and Julian Berberich and Frank Allgöwer}, doi = {10.1016/j.automatica.2025.112612}, journal = {Automatica}, month = {1}, pages = {112612}, shortjournal = {Automatica}, title = {Data-driven min–max MPC for linear systems: Robustness and adaptation}, volume = {183}, year = {2026}, }
• (2026). Local practically safe extremum seeking with assignable rate of attractivity to the safe set. AUTOM, 183, 112611. (www)

  @article{ , abstract = {We present Assignably Safe Extremum Seeking (ASfES), an algorithm designed to minimize a measured, static objective function while maintaining a measured, static metric of safety (a control barrier function or CBF) to be positive in a practical sense. We ensure that for trajectories with safe initial conditions, the violation of safety can be made arbitrarily small through appropriately chosen design constants. We also guarantee an assignable “attractivity” rate: from unsafe initial conditions, the trajectories approach the safe set, in the sense of the measured CBF, at a rate no slower than a user-assigned rate. Similarly, from safe initial conditions, the trajectories approach the unsafe set, in the sense of the CBF, no faster than the assigned attractivity rate. The feature of assignable attractivity is not present in the semiglobal version of safe extremum seeking, where the semiglobality of convergence is achieved by slowing the adaptation. We also demonstrate local convergence of the parameter to a neighborhood of the minimum of a quadratic objective function constrained to the safe set with a linear CBF. The ASfES algorithm and analysis are multivariable, but we also extend the algorithm to a Newton-Based ASfES scheme which we show is only useful in the scalar case. The proven properties of the designs are illustrated through simulation examples.}, archive = {J_AUTOM}, author = {Alan Williams and Miroslav Krstic and Alexander Scheinker}, doi = {10.1016/j.automatica.2025.112611}, journal = {Automatica}, month = {1}, pages = {112611}, shortjournal = {Automatica}, title = {Local practically safe extremum seeking with assignable rate of attractivity to the safe set}, volume = {183}, year = {2026}, }
• (2026). Secondary safety control for systems with sector bounded nonlinearities. AUTOM, 183, 112610. (www)

  @article{ , abstract = {We consider the problem of safety verification and safety-aware controller synthesis for systems with sector bounded nonlinearities. We aim to keep the states of the system within a given safe set under potential actuator and sensor attacks. Specifically, we adopt the setup that a controller has already been designed to stabilize the plant. Using invariant sets and barrier certificate theory, we first give sufficient conditions to verify the safety of the closed-loop system under attacks. Furthermore, by using a subset of sensors that are assumed to be free of attacks, we provide a synthesis method for a secondary controller that enhances the safety of the system. The sufficient conditions to verify safety are derived using Lyapunov-based tools and the S -procedure. Using the projection lemma, the conditions are then formulated as linear matrix inequality (LMI) problems which can be solved efficiently. Lastly, our theoretical results are illustrated through numerical simulations.}, archive = {J_AUTOM}, author = {Yankai Lin and Michelle S. Chong and Carlos Murguia}, doi = {10.1016/j.automatica.2025.112610}, journal = {Automatica}, month = {1}, pages = {112610}, shortjournal = {Automatica}, title = {Secondary safety control for systems with sector bounded nonlinearities}, volume = {183}, year = {2026}, }
• (2026). Optimal control of continuous-time symmetric systems with unknown dynamics and noisy measurements. AUTOM, 183, 112609. (www)

  @article{ , abstract = {An iterative learning algorithm is presented for continuous-time linear–quadratic optimal control problems where the system is externally symmetric with unknown dynamics. Both finite-horizon and infinite-horizon problems are considered. It is shown that the proposed algorithm is globally convergent to the optimal solution and has some advantages over adaptive dynamic programming, including unbiased performance under noisy measurements, relatively low computational burden, and no requirement for exploration noise. Numerical experiments show the effectiveness of the results.}, archive = {J_AUTOM}, author = {Hamed Taghavian and Florian Dörfler and Mikael Johansson}, doi = {10.1016/j.automatica.2025.112609}, journal = {Automatica}, month = {1}, pages = {112609}, shortjournal = {Automatica}, title = {Optimal control of continuous-time symmetric systems with unknown dynamics and noisy measurements}, volume = {183}, year = {2026}, }
• (2026). Non-parametric IQC multipliers in data-driven robust controller synthesis. AUTOM, 183, 112608. (www)

  @article{ , abstract = {The paper presents a robust data-driven controller synthesis method for generalised multi-input multi-output (MIMO) systems. Using the frequency response of a linear time-invariant (LTI) MIMO system and characterising perturbations through Integral Quadratic Constraint (IQC), the method provides a convex set of controllers robust to perturbations. This facilitates the design of controllers with either robust stability or robust performance criteria. Notably, the proposed method is versatile, as it is also applicable for non-parametric IQC multipliers. An example of a non-parametric IQC multiplier for elliptical uncertainty quantification is demonstrated and subsequently employed in designing a robust controller for a hybrid active-passive micro-vibration platform. Experimental results show that the synthesised controller effectively achieves the desired levels of both robustness and performance.}, archive = {J_AUTOM}, author = {Vaibhav Gupta and Elias Klauser and Alireza Karimi}, doi = {10.1016/j.automatica.2025.112608}, journal = {Automatica}, month = {1}, pages = {112608}, shortjournal = {Automatica}, title = {Non-parametric IQC multipliers in data-driven robust controller synthesis}, volume = {183}, year = {2026}, }
• (2026). Iterative learning control for performance-driven switched systems under all unknown channel gains. AUTOM, 183, 112607. (www)

  @article{ , abstract = {Fading channels introduce random and unknown gains into signal transmission, which has a great impact on the accuracy of the transmitted data, thus affecting system stability. However, there are few studies on this regard in the control field. In this paper, the transmission of both feedback and switching signals is affected by the unknown channel gains, which induces severe mode asynchronization in the switched systems under iterative learning control (ILC). For the faded switching signals, by designing a calibration-rounding mechanism (CRM) to recover it, the asynchronous time can gradually decrease with the iteration, and finally the system achieves synchronization, that is calibration-rounding-based iterative synchronization (CRIS). For the faded feedback signals, an improved gain estimation mechanism (IGEM) is given to correct the signal faster with fewer iterations. Moreover, a novel tracking performance-driven switching law (TPD-SL) is proposed to reasonably schedule subsystems for obtaining optimal performance. A set of new average dwell time (ADT) conditions with the minimum synchronization time is obtained. Finally, the effectiveness of the method is verified by numerical simulations.}, archive = {J_AUTOM}, author = {Yiwen Qi and Ziyu Qu and Dong Shen}, doi = {10.1016/j.automatica.2025.112607}, journal = {Automatica}, month = {1}, pages = {112607}, shortjournal = {Automatica}, title = {Iterative learning control for performance-driven switched systems under all unknown channel gains}, volume = {183}, year = {2026}, }
• (2026). Induced norm analysis of linear systems for nonnegative input signals. AUTOM, 183, 112606. (www)

  @article{ , abstract = {This paper is concerned with the analysis of the L p ( p ∈ [ 1 , ∞ ) , p = ∞ ) induced norms of continuous-time linear systems where input signals are restricted to be nonnegative. This norm is referred to as the L p + induced norm in this paper. It has been shown recently that the L 2 + induced norm is effective for the stability analysis of nonlinear feedback systems where the nonlinearity returns only nonnegative signals. However, the exact computation of the L 2 + induced norm is essentially difficult. To get around this difficulty, in the first part of this paper, we provide a copositive-programming-based method for the upper bound computation by capturing the nonnegativity of the input signals by copositive multipliers. In the second part, we consider how far the L 2 + induced norm can be smaller than the standard L 2 induced norm, and derive the uniform infimum on the ratio of the L 2 + induced norm to the L 2 induced norm over all linear systems including infinite-dimensional ones. Then, for each linear system, we finally derive a computation method of the lower bounds of the L 2 + induced norm that are larger than (or equal to) the value determined by the uniform infimum. The effectiveness of the upper/lower bound computation methods is illustrated by numerical examples.}, archive = {J_AUTOM}, author = {Yoshio Ebihara and Noboru Sebe and Hayato Waki and Dimitri Peaucelle and Sophie Tarbouriech and Victor Magron and Tomomichi Hagiwara}, doi = {10.1016/j.automatica.2025.112606}, journal = {Automatica}, month = {1}, pages = {112606}, shortjournal = {Automatica}, title = {Induced norm analysis of linear systems for nonnegative input signals}, volume = {183}, year = {2026}, }
• (2026). Distributed fault-tolerant control for a class of uncertain cascaded ODE-PDE multi-agent systems. AUTOM, 183, 112605. (www)

  @article{ , abstract = {In this paper, a distributed fault-tolerant control is proposed for a class of uncertain cascaded ordinary differential equation (ODE)-partial differential equation (PDE) multi-agents. For the cascaded system under consideration, the output of ODEs serves as the input signal of PDEs such that a distributed control involving states of the ODE as well as the PDE system is expected to be designed for consensus tracking. Based on the designed distributed control, where the actuator is assumed to be healthy, a distributed fault-tolerant control is designed to guarantee the performance of the system, and it is further improved to deal with actuator failures. To enhance the robustness of the cascaded system, the system uncertainties and the external disturbances of both followers and leader are compensated using adaptive laws and an adaptive distributed observer, based on which the stabilization of the cascaded ODE-PDE multi-agent system can be achieved with partially known agents’ information. Simulation results are carried out to prove the effectiveness of the proposed method.}, archive = {J_AUTOM}, author = {Xueyan Xing and Guoqiang Hu}, doi = {10.1016/j.automatica.2025.112605}, journal = {Automatica}, month = {1}, pages = {112605}, shortjournal = {Automatica}, title = {Distributed fault-tolerant control for a class of uncertain cascaded ODE-PDE multi-agent systems}, volume = {183}, year = {2026}, }
• (2026). A compensation-oriented algorithm for difference-driven identification under binary-valued observations and data packet dropout. AUTOM, 183, 112604. (www)

  @article{ , abstract = {This paper investigates the identification problem for finite impulse response (FIR) systems with binary-valued observations under event-triggered communication mechanism and data packet dropout. The challenge lies in the inability to distinguish between untriggered events and packet loss when no information is received, which prevents us from obtaining the statistical properties of the binary-valued sequence. A compensation-oriented difference-driven identification (CODD) algorithm is proposed to estimate the parameter by recovering the mean of the original binary-valued sequence, where different values for the observation estimates are assigned when receiving 0, 1 or nothing. Even though, the convergence analysis of the parameter estimate is still challenging since the assigned values are dependent. To tackle this difficulty, the estimate error is divided into two parts: an initial assigned value related part, which is demonstrated to be convergent through the construction of an auxiliary set, and the remaining component, which happens to be a convergent martingale-difference sequence. As a result, the almost sure convergence and the asymptotic normality of the CODD algorithm are established when data packet loss probability is less than 1 2 . By calculating the communication rate, it is proven that the difference-driven mechanism can save 50% of the communication cost compared to original binary-valued systems. Furthermore, when data packet loss probability is high, an m -channel compensation-oriented identification ( m -CODD) algorithm is constructed by utilizing retransmission of the each observation for m times, which is designed based on the packet loss probability. The properties of m -CODD algorithm including convergence, asymptotic normality and communication rate are established. Numerical simulations are illustrated to show the theoretical results.}, archive = {J_AUTOM}, author = {Tianning Han and Ying Wang and Jin Guo and Yanlong Zhao}, doi = {10.1016/j.automatica.2025.112604}, journal = {Automatica}, month = {1}, pages = {112604}, shortjournal = {Automatica}, title = {A compensation-oriented algorithm for difference-driven identification under binary-valued observations and data packet dropout}, volume = {183}, year = {2026}, }
• (2026). Distributed nash equilibrium seeking for aggregative games of linear systems subject to unknown disturbances. AUTOM, 183, 112603. (www)

  @article{ , abstract = {In this paper, we address the distributed Nash equilibrium seeking problem for aggregative games of N players subject to unknown disturbances over strongly connected networks. Compared with existing works, the general linear dynamics, general directed and strongly connected networks, as well as unknown disturbances are tackled simultaneously in the aggregative games. First, by introducing certain coordinate transformation and feedback linearization method, we develop a distributed gradient-based Nash equilibrium seeking law. A dynamic average consensus dynamics is designed to deal with the challenge by unbalance of general strongly connected networks. By the graph-related property and converse Lyapunov theorem, we establish the global exponential stability of a linear system and a class of nonlinear systems, respectively. Then, we propose a gain design method to obtain the stability of the nonlinear closed-loop system, which is not in the lower triangular form. Inspired by the output regulation theory, we design an internal model and an adaptive dynamics to tackle the unknown disturbances. Resorting to the perturbation theory and the internal model principle, we demonstrate that distributed Nash equilibrium seeking for aggregative games of N players with general linear systems subject to unknown disturbances over strongly connected networks can be achieved. Finally, the effectiveness of the proposed distributed Nash equilibrium seeking approaches are verified by their applications to some simulation examples.}, archive = {J_AUTOM}, author = {Lupeng Liu and Fang Deng and Jie Chen and Maobin Lu}, doi = {10.1016/j.automatica.2025.112603}, journal = {Automatica}, month = {1}, pages = {112603}, shortjournal = {Automatica}, title = {Distributed nash equilibrium seeking for aggregative games of linear systems subject to unknown disturbances}, volume = {183}, year = {2026}, }
• (2026). Adaptive fault-tolerant control of nonlinear systems: A self-regulating spatiotemporal performance approach. AUTOM, 183, 112602. (www)

  @article{ , abstract = {This paper investigates the problem of performance constraints of uncertain nonlinear systems with actuator faults. Particularly, by developing the boundary threshold triggering strategy, the performance boundaries would be updated adaptively once the distance between the tracking error and the performance boundaries is smaller than the given threshold. Moreover, compared with the existing works, the transient behaviors are improved, and the limitation imposed on the initial conditions is removed attributed to the construction of novel performance functions and error transformation. Then, an adaptive fault-tolerant control scheme with self-regulating spatiotemporal performance is designed for a class of nonlinear systems with non-parametric uncertainties. It is shown that both the boundedness of the closed-loop signals and the satisfactory performance constraints are guaranteed in the presence of unpredictable actuator failure. The effectiveness of the proposed method is verified by theoretical analysis and numerical simulation.}, archive = {J_AUTOM}, author = {Zeqiang Li and Jason J.R. Liu and Yongduan Song}, doi = {10.1016/j.automatica.2025.112602}, journal = {Automatica}, month = {1}, pages = {112602}, shortjournal = {Automatica}, title = {Adaptive fault-tolerant control of nonlinear systems: A self-regulating spatiotemporal performance approach}, volume = {183}, year = {2026}, }
• (2026). Fully coupled nonlinear FBS△Es: Solvability and LQ control insights. AUTOM, 183, 112601. (www)

  @article{ , abstract = {In this paper, a class of fully coupled nonlinear forward–backward stochastic difference equations (FBS △ Es) is proposed and the existence of solutions is proved based on a linear-quadratic (LQ) optimal control problem. Inspired from the solvability studies of various forward–backward stochastic differential equations (FBSDEs), the dominant-monotone framework is discretized and a continuum approach is used to prove the unique solvability of the fully coupled FBS △ Es and to obtain a pair of estimates on the solutions, and finally, the conclusions are applied to the related LQ problem.}, archive = {J_AUTOM}, author = {Zhipeng Niu and Qingxin Meng and Xun Li and Maoning Tang}, doi = {10.1016/j.automatica.2025.112601}, journal = {Automatica}, month = {1}, pages = {112601}, shortjournal = {Automatica}, title = {Fully coupled nonlinear FBS△Es: Solvability and LQ control insights}, volume = {183}, year = {2026}, }
• (2026). Aerial target encirclement and interception with noisy range observations. AUTOM, 183, 112600. (www)

  @article{ , abstract = {This paper proposes a strategy to encircle and intercept a non-cooperative aerial point-mass moving target by leveraging noisy range measurements for state estimation. In this approach, the guardians actively ensure the observability of the target by using an anti-synchronization (AS), 3D “vibrating string” trajectory, which enables rapid position and velocity estimation based on the Kalman filter. Additionally, a novel anti-target controller is designed for the guardians to enable adaptive transitions from encircling a protected target to encircling, intercepting, and neutralizing a hostile target, taking into consideration the input constraints of the guardians. Based on the guaranteed uniform observability, the exponentially bounded stability of the state estimation error and the convergence of the encirclement error are rigorously analyzed. Simulation results and real-world UAV experiments are presented to further validate the effectiveness of the system design.}, archive = {J_AUTOM}, author = {Fen Liu and Shenghai Yuan and Thien-Minh Nguyen and Wei Meng and Lihua Xie}, doi = {10.1016/j.automatica.2025.112600}, journal = {Automatica}, month = {1}, pages = {112600}, shortjournal = {Automatica}, title = {Aerial target encirclement and interception with noisy range observations}, volume = {183}, year = {2026}, }
• (2026). Stochastic gradient hamiltonian sequential monte carlo filter with earth mover’s distance sampling for target tracking. AUTOM, 183, 112599. (www)

  @article{ , abstract = {High-speed missile tracking requires advanced filtering algorithms to overcome the limitations of traditional nonlinear methods. This study presents the stochastic gradient Hamiltonian sequential Monte Carlo filter, combining stochastic gradient Hamilton Monte Carlo with sequential Monte Carlo (SMC) for enhanced sampling performance and reduced computational burden. The method incorporates Earth Mover’s Distance based adaptive resampling with theoretical bounds for optimal particle count. Validation through univariate nonstationary growth model simulation and bearing-only tracking experiments demonstrates superior performance over conventional methods, achieving 15 % root mean square error improvement compared to conventional SMC and 30 % over extended Kalman filter/unscented Kalman filter approaches.}, archive = {J_AUTOM}, author = {Chang Ho Kang and Sun Young Kim}, doi = {10.1016/j.automatica.2025.112599}, journal = {Automatica}, month = {1}, pages = {112599}, shortjournal = {Automatica}, title = {Stochastic gradient hamiltonian sequential monte carlo filter with earth mover’s distance sampling for target tracking}, volume = {183}, year = {2026}, }
• (2026). Distributed nash equilibrium seeking with a dynamic set of players. AUTOM, 183, 112598. (www)

  @article{ , abstract = {This paper formulates a new distributed Nash equilibrium seeking problem with a dynamic set of players in which players are allowed to join and leave the network in a free manner during the decision-making process. To accommodate the dynamic joining and leaving behaviors of the players, a status estimation mechanism, which is capable of estimating in a finite time whether the players are active or inactive, is introduced. Based on the status estimation mechanism, a gradient play based algorithm is developed for distributed Nash equilibrium seeking in the dynamic environment. It is shown that under strongly connected communication graphs, players’ actions are convergent to a small neighborhood of the new Nash equilibrium linearly every time the player set changes. Moreover, the convergence accuracy and convergence rate can be adjusted by suitably tuning the step-size. To cover more general communication scenarios, strongly connected graphs are further relaxed to be B-jointly connected graphs, under which the convergence properties of the proposed algorithm are analytically studied. Furthermore, the upper bound of the average tracking error is quantified to evaluate the dynamic performance of the proposed algorithm. In the last, a simulation study on energy consumption games is given to verify the effectiveness of the proposed algorithm.}, archive = {J_AUTOM}, author = {Yuxuan Liu and Maojiao Ye and Lei Ding and Lihua Xie and Qing-Long Han}, doi = {10.1016/j.automatica.2025.112598}, journal = {Automatica}, month = {1}, pages = {112598}, shortjournal = {Automatica}, title = {Distributed nash equilibrium seeking with a dynamic set of players}, volume = {183}, year = {2026}, }
• (2026). Constrained finite-time and fixed-time stabilization for linear systems: Adaptive implicit lyapunov function-based control. AUTOM, 183, 112597. (www)

  @article{ , abstract = {In this paper, finite-time and fixed-time stabilization problems are investigated for single-input single-output (SISO) linear system under the control input constraint. The achievement of finite-time or fixed-time convergence rates is facilitated through the utilization of adaptive implicit Lyapunov function (ILF)-based control. For ease of practical implementation, the dynamics of Approximated-ILF (AILF) guarantees the precise estimation of ILF, while the stability of AILF-based control holds established. Furthermore, from both performance and input-constrained safety considerations, the anti-windup (AW) AILF endows the system with tolerance to saturation and maintains the non-asymptotic convergence properties. Numerical simulations support the obtained theoretical results and verify their effectiveness.}, archive = {J_AUTOM}, author = {Peng Wang and Mou Chen and Shuzhi Sam Ge and Xiaobing Zhang}, doi = {10.1016/j.automatica.2025.112597}, journal = {Automatica}, month = {1}, pages = {112597}, shortjournal = {Automatica}, title = {Constrained finite-time and fixed-time stabilization for linear systems: Adaptive implicit lyapunov function-based control}, volume = {183}, year = {2026}, }
• (2026). Input-to-state stability of self-triggered impulsive control systems. AUTOM, 183, 112596. (www)

  @article{ , abstract = {This paper studies the local input-to-state stability ( LISS ) of nonlinear systems under the self-triggered impulsive control ( STIC ) method. A novel LISS -type comparison principle is proposed, by which the LISS property of addressed system can be derived by the LISS property of its comparison system. On the basis of it, some Lyapunov-based sufficient conditions for non-Zeno behavior and LISS of nonlinear systems are provided in the framework of STIC . Moreover, the designed self-triggering mechanism ( STM ) is in the form of an explicit relationship with simple structure and east implementation. Finally, two numerical examples are given to illustrate the effectiveness of the proposed results.}, archive = {J_AUTOM}, author = {Xiaodi Li and Mingzhu Wang}, doi = {10.1016/j.automatica.2025.112596}, journal = {Automatica}, month = {1}, pages = {112596}, shortjournal = {Automatica}, title = {Input-to-state stability of self-triggered impulsive control systems}, volume = {183}, year = {2026}, }
• (2026). Event-triggered global and group consensus for linear multiagent systems with multiplicative noises. AUTOM, 183, 112595. (www)

  @article{ , abstract = {This article is concerned with the event-triggered global and group consensus problems for linear multiagent systems (MASs) with multiplicative noises and directed topology. First, for the global consensus of MASs under a balanced directed graph, a stochastic input-to-state stability property is proved to show that the piecewise constant control input has sufficient feedback capability; based on the relative state measurement, two novel static and dynamic event-triggered mechanisms (ETMs) are established to reduce the number of controller updates, where the Zeno phenomenon is effectively eliminated by imposing a uniformly strictly fixed positive time for all the inter-event times. Especially, new stochastic stability analysis techniques are developed to give sufficient conditions of mean square and almost sure consensus in both static and dynamic ETMs. Moreover, the obtained results on global consensus are extended to the setting of group consensus in order to cope with some grouping tasks in practical applications. Finally, some examples are presented to show the validity of theoretical results.}, archive = {J_AUTOM}, author = {Ruru Jia and Yuan-Hua Ni and Guangchen Wang}, doi = {10.1016/j.automatica.2025.112595}, journal = {Automatica}, month = {1}, pages = {112595}, shortjournal = {Automatica}, title = {Event-triggered global and group consensus for linear multiagent systems with multiplicative noises}, volume = {183}, year = {2026}, }
• (2026). Data-model-hybrid-driven near-optimal operational control of two-time-scale industrial systems with unknown operational model. AUTOM, 183, 112594. (www)

  @article{ , abstract = {This article discusses the optimal operational control (OOC) problem of two-time-scale (TTS) industrial systems with unknown operation model. Based on the singular perturbation theory (SPT), the OOC problem of TTS industrial systems is decomposed into an optimal regulation problem in fast time-scale and an optimal set-point tracking problem in slow time-scale. Then, by convex duality, the obtained optimization problems are equivalently transformed into convex optimization (CO) problems, and a data-model-hybrid-driven composite controller is designed. The design method of this composite controller avoids the potential numerical stiffness problems, and does not need complete system dynamics information while ensuring the steady-state output tracking error converges to zero. Finally, an example of mixed separation thickening process (MSTP) of hematite beneficiation is given to show the effectiveness of the proposed scheme.}, archive = {J_AUTOM}, author = {Yao Xu and Linna Zhou and Jianguo Zhao and Lei Ma and Chunyu Yang}, doi = {10.1016/j.automatica.2025.112594}, journal = {Automatica}, month = {1}, pages = {112594}, shortjournal = {Automatica}, title = {Data-model-hybrid-driven near-optimal operational control of two-time-scale industrial systems with unknown operational model}, volume = {183}, year = {2026}, }
• (2026). Event-triggered distributed optimal coordination of heterogeneous linear MASs over weight-unbalanced digraphs. AUTOM, 183, 112593. (www)

  @article{ , abstract = {In this paper, the distributed optimal coordination problem of heterogeneous linear multi-agent systems is studied, in which the actual information of the left eigenvector and out-degree cannot be obtained in advance. The underlying communication network among agents is assumed to be a weight-unbalanced and strongly connected digraph. The global objective function of the whole system is the sum of privately-known local objective functions, which are assumed to be strongly convex with Lipschitz continuous gradients. To achieve the distributed optimal coordination with low communication and computational resource consumption, and to efficiently reduce the frequency of control signal updates, a new distributed event-triggered control scheme is proposed. In this scheme, an auxiliary system with adaptive edge weights is established to generate the optimal trajectories and the tracking control inputs are designed for agents to track the given optimal trajectories. It is privacy-preserving to some extent as no actual state or output information of agents is communicated with neighbors. Sufficient conditions are derived, under which the distributed optimal coordination under consideration can be exponentially achieved. Moreover, the Zeno-behavior is shown to be excluded. Additionally, for cases where agent states are unmeasurable, a new distributed event-triggered observer-based control input is designed for each agent to achieve optimal coordination. Finally, numerical simulations are carried out to show the good performance of the proposed control scheme.}, archive = {J_AUTOM}, author = {Dandan Wang and Jialing Zhou and Guanghui Wen}, doi = {10.1016/j.automatica.2025.112593}, journal = {Automatica}, month = {1}, pages = {112593}, shortjournal = {Automatica}, title = {Event-triggered distributed optimal coordination of heterogeneous linear MASs over weight-unbalanced digraphs}, volume = {183}, year = {2026}, }
• (2026). Reducing real-time complexity via sub-control lyapunov functions: From theory to experiments. AUTOM, 183, 112592. (www)

  @article{ , abstract = {The techniques to design control Lyapunov functions (CLF), along with a proper stabilizing feedback, possibly in the presence of constraints, often provide control laws that are too complex for proper implementation online, especially when an optimization problem is involved. In this work, we show how to acquire an alternative, computationally attractive feedback. Given a nominal CLF and a nominal state feedback, we say that a different positive definite function is a Sub-control Lyapunov function (SCLF) if its Lyapunov derivative is negative-definite and bounded above by the Lyapunov derivative of the nominal function with the nominal control. It turns out that if we consider a family of basis functions, then an SCLF can be computed by linear programming, with an infinite number of constraints. The idea is that although the offline computational burden to achieve the new controller and solve the linear program is considerable, the online computational burden is drastically reduced. Comprehensive simulations and experiments on drone control are conducted to demonstrate the effectiveness of the study.}, archive = {J_AUTOM}, author = {Huu-Thinh Do and Franco Blanchini and Stefano Miani and Ionela Prodan}, doi = {10.1016/j.automatica.2025.112592}, journal = {Automatica}, month = {1}, pages = {112592}, shortjournal = {Automatica}, title = {Reducing real-time complexity via sub-control lyapunov functions: From theory to experiments}, volume = {183}, year = {2026}, }
• (2026). Adversarial dynamic games for markov jump systems: A policy iteration Q-learning method. AUTOM, 183, 112591. (www)

  @article{ , abstract = {This paper presents a reinforcement Q-learning approach for solving adversarial dynamic games in Markov jump systems. The H ∞ control problem is first formulated as a two-player zero-sum dynamic game, where the control policy and the disturbance policy act as adversarial players. To derive the Nash equilibrium control strategies for such games, a set of coupled algebraic Riccati equations is established, with the disturbance attenuation level properly prescribed. On this basis, two novel data-driven parallel Q-learning algorithms are proposed. The advantages of the proposed method are threefold: (i) it does not require precise knowledge of the system dynamics; (ii) it learns the optimal disturbance attenuation level; (iii) it yields Nash equilibrium control strategies. Finally, two simulation examples validate the effectiveness of the proposed method.}, archive = {J_AUTOM}, author = {Hao Shen and Jiacheng Wu and Jing Wang and Zhengguang Wu}, doi = {10.1016/j.automatica.2025.112591}, journal = {Automatica}, month = {1}, pages = {112591}, shortjournal = {Automatica}, title = {Adversarial dynamic games for markov jump systems: A policy iteration Q-learning method}, volume = {183}, year = {2026}, }
• (2026). H2/H∞ state and output feedback control with sparse actuation. AUTOM, 183, 112581. (www)

  @article{ , abstract = {In this paper, we present novel convex optimization formulations for designing full-state and output-feedback controllers with sparse actuation that achieve user-specified H 2 and H ∞ performance criteria. The sparsity is induced through the ℓ 1 -minimization over channel-wise H 2 norms from disturbances to the individual actuator signals, while simultaneously constraining H 2 or H ∞ norm from disturbances to the output variables The proposed approach is applied to a structural dynamics problem, demonstrating the advantages of simultaneous optimization of the control law and the actuation architecture in realizing an efficient closed-loop system, as well as highlighting the trade-offs between maximum allowable actuator magnitudes and the controller sparsity.}, archive = {J_AUTOM}, author = {Vedang M. Deshpande and Raktim Bhattacharya}, doi = {10.1016/j.automatica.2025.112581}, journal = {Automatica}, month = {1}, pages = {112581}, shortjournal = {Automatica}, title = {H2/H∞ state and output feedback control with sparse actuation}, volume = {183}, year = {2026}, }
• (2026). Adaptive dynamic event–triggered distributed optimal coordination of heterogeneous uncertain nonlinear multiagent systems. AUTOM, 183, 112580. (www)

  @article{ , abstract = {This paper addresses the distributed optimal coordination problem for a class of heterogeneous uncertain nonlinear multiagent systems. Instead of relying on the analytical forms of gradient functions, we use the measured gradient values depending on agents’ real-time outputs and propose a novel adaptive distributed control scheme. This scheme integrates event-triggered optimal coordinators, high-order filters, and tracking controllers. To handle the interaction between optimal coordinators and filters, we incorporate a new compensation term into the updating law for the coupling weight of each edge. Moreover, we design a novel adaptive distributed dynamic event-triggering mechanism that ensures that the inter-event times of each agent are lower bounded by a positive constant. Asymptotic convergence of agents’ outputs to the optimal point is proved by constructing a composite Lyapunov function. The proposed control scheme does not depend on global topology information. A numerical example is given to demonstrate the effectiveness of the proposed control scheme.}, archive = {J_AUTOM}, author = {Tianyu Liu and Lu Liu}, doi = {10.1016/j.automatica.2025.112580}, journal = {Automatica}, month = {1}, pages = {112580}, shortjournal = {Automatica}, title = {Adaptive dynamic event–triggered distributed optimal coordination of heterogeneous uncertain nonlinear multiagent systems}, volume = {183}, year = {2026}, }
• (2026). Accelerated primal–dual methods for strongly convex objective functions in continuous and discrete time. AUTOM, 183, 112579. (www)

  @article{ , abstract = {In this paper, we introduce a “second-order primal” + “first-order dual” continuous-time dynamic for linearly constrained optimization problems, where the objective function is μ -strongly convex. We consider a constant damping 2 μ for the second-order ordinary differential equation in the primal variable, following Nesterov’s acceleration for strongly convex optimization. A positive constant scaling is applied to the primal variable, while a positive increasing scaling function is applied to the dual variable. We prove that the proposed dynamic achieves a fast convergence rate for both the objective residual and the feasibility violation, with the decay rate potentially reaching O ( e − μ t ) . Additionally, we show that the dynamic is robust under small perturbations. By discretizing the proposed continuous-time dynamic, we develop an accelerated linearized augmented Lagrangian method for strongly convex composite optimization with linear constraints, where the objective function has a nonsmooth + smooth composite structure. The proposed algorithm achieves a fast convergence rate that matches the one of the continuous-time dynamic. We also consider an inexact version of the proposed algorithm, which can be viewed as a discrete version of the perturbed continuous-time dynamic. Numerical results are provided to verify the practical performances.}, archive = {J_AUTOM}, author = {Xin He and Dong He and Ya-Ping Fang}, doi = {10.1016/j.automatica.2025.112579}, journal = {Automatica}, month = {1}, pages = {112579}, shortjournal = {Automatica}, title = {Accelerated primal–dual methods for strongly convex objective functions in continuous and discrete time}, volume = {183}, year = {2026}, }
• (2026). Hierarchical supervisory control of networked and cyber-attacked discrete-event systems. AUTOM, 183, 112578. (www)

  @article{ , abstract = {In standard supervisory control of discrete-event systems, partial (incomplete) observations are given by deterministic functions such as natural projections, which erase unobservable events, or masks, which can represent indistinguishable events, where two or more different events yield the same observation. However, communication channels in modern technological systems are not always reliable and can be attacked by malicious external agents. In that case, the plant observations obtained by the supervisor may not be deterministic, e.g., due to delays and losses, or external attacks. This paper considers a unified supervisory control framework with set-valued (nondeterministic) observations and proposes a simplified version of nondeterministic observability, together with a generalized normality. It shows how the results of hierarchical control can be extended to the networked and cyber-attacked discrete-event systems at the same time.}, archive = {J_AUTOM}, author = {Shaowen Miao and Jan Komenda and Feng Lin}, doi = {10.1016/j.automatica.2025.112578}, journal = {Automatica}, month = {1}, pages = {112578}, shortjournal = {Automatica}, title = {Hierarchical supervisory control of networked and cyber-attacked discrete-event systems}, volume = {183}, year = {2026}, }
• (2026). On the design of linear time varying model predictive control for trajectory stabilization. AUTOM, 183, 112577. (www)

  @article{ , abstract = {Stabilizing a reference trajectory of a nonlinear system is a recurrent, non-trivial task in control engineering. A common approach is to linearize the dynamics along the trajectory, thus deriving a linear-time-varying (LTV) model, and to design a model predictive controller (MPC), which results to be computationally efficient, since only convex programs need to be solved in real time, while retaining constraint handling capabilities. Building on recent developments in gain-scheduling control design, where linearization errors and tracking error bounds are considered, a new approach to derive such LTV-MPC controllers is presented. The method addresses the systematic derivation of a suitable terminal cost. The resulting MPC law is tube-based, exploiting the co-designed auxiliary gain-scheduled controller. Computational and implementation aspects are discussed as well, and the resulting hierarchical method is demonstrated both in simulation and in experiments with a small drone with fast dynamics and limited embedded computational capacity.}, archive = {J_AUTOM}, author = {Nicolas Kessler and Lorenzo Fagiano}, doi = {10.1016/j.automatica.2025.112577}, journal = {Automatica}, month = {1}, pages = {112577}, shortjournal = {Automatica}, title = {On the design of linear time varying model predictive control for trajectory stabilization}, volume = {183}, year = {2026}, }
• (2026). Intersection-based architectures for decentralized diagnosis of discrete event systems. AUTOM, 183, 112576. (www)

  @article{ , abstract = {In this paper, two intersection-based architectures, named the normal-state-estimator-intersection-based architecture (N-SEI architecture) and the failure-state-estimator-intersection-based architecture (F-SEI architecture), are examined for decentralized diagnosis of discrete event systems. For each of these architectures, the corresponding notion of codiagnosability is defined. These defined notions of codiagnosability are incomparable with inference diagnosability for the inference-based architecture. In addition, codiagnosability for the N-SEI architecture is weaker than the existing notion of intersection-based codiagnosability, while codiagnosability for the F-SEI architecture is incomparable with it. For each of the N-SEI and F-SEI architectures, a method for verifying the corresponding notion of codiagnosability is developed.}, archive = {J_AUTOM}, author = {Shigemasa Takai and Takashi Yamamoto}, doi = {10.1016/j.automatica.2025.112576}, journal = {Automatica}, month = {1}, pages = {112576}, shortjournal = {Automatica}, title = {Intersection-based architectures for decentralized diagnosis of discrete event systems}, volume = {183}, year = {2026}, }
• (2026). Distributed stochastic constrained optimization with constant step-sizes via saddle-point dynamics. AUTOM, 183, 112575. (www)

  @article{ , abstract = {This paper considers distributed stochastic optimization problems over a multi-agent network, where each agent collaboratively minimizes the sum of individual expectation-valued cost functions subject to nonidentical set constraints. We first recast the distributed constrained optimization as a constrained saddle-point problem. Subsequently, two distributed stochastic algorithms via optimistic gradient descent ascent (SOGDA) and extragradient (SEG) methods are developed with constant step sizes, in which the variable sample-size technique is incorporated to reduce the variance of the sampled gradients. We present the explicit selection criteria of the constant step size, under which the developed algorithms achieve almost sure convergence to an optimal solution. Moreover, the convergence rate is O ( 1 / k ) for merely convex cost functions, which matches the optimal rate of its deterministic counterpart. Finally, a numerical example is provided to reflect the theoretical findings.}, archive = {J_AUTOM}, author = {Yi Huang and Shisheng Cui and Xianlin Zeng and Ziyang Meng}, doi = {10.1016/j.automatica.2025.112575}, journal = {Automatica}, month = {1}, pages = {112575}, shortjournal = {Automatica}, title = {Distributed stochastic constrained optimization with constant step-sizes via saddle-point dynamics}, volume = {183}, year = {2026}, }
• (2026). A characterization method of terminal ingredients for nonlinear MPC using value-based reinforcement learning. AUTOM, 183, 112574. (www)

  @article{ , abstract = {The stability of nonlinear model predictive control (MPC) relies significantly on stabilizing factors such as the terminal region and cost. A larger terminal region not only expands the region of attraction for the closed-loop system but also contributes to reducing online computation costs. However, existing methods in the literature often impose limitations on the degrees of freedom available for characterizing terminal ingredients. This limitation arises from the reliance on either a predetermined linear local controller or a preset control Lyapunov function. This paper introduces an innovative approach to terminal ingredient characterization leveraging value-based reinforcement learning (RL). This method provides ample degrees of freedom for expanding the terminal region. To achieve this, a deep neural network is employed to learn the parametric state value function, serving as the terminal cost for MPC. The local controller adopts a one-step MPC instead of a predetermined linear or nonlinear feedback controller. Subsequently, a terminal set sequence is constructed iteratively through the one-step set expansion. The proposed approach’s effectiveness is validated through simulations.}, archive = {J_AUTOM}, author = {Jinghan Cui and Jinwu Gao and Xiangjie Liu and Yuqi Liu and Shuyou Yu}, doi = {10.1016/j.automatica.2025.112574}, journal = {Automatica}, month = {1}, pages = {112574}, shortjournal = {Automatica}, title = {A characterization method of terminal ingredients for nonlinear MPC using value-based reinforcement learning}, volume = {183}, year = {2026}, }
• (2026). Aperiodic-sampled neural network controllers with closed-loop stability verifications. AUTOM, 183, 112573. (www)

  @article{ , abstract = {In this paper, we synthesize two aperiodic-sampled deep neural network (DNN) control schemes, based on the closed-loop tracking stability guarantees. By means of the integral quadratic constraint coping with the input–output behavior of system uncertainties/nonlinearities and the convex relaxations of nonlinear DNN activations leveraging their local sector-bounded attributes, we establish conditions to design the event- and self-triggered logics and to compute the ellipsoidal inner approximations of region of attraction, respectively. Finally, we perform a numerical example of an inverted pendulum to illustrate the effectiveness of the proposed aperiodic-sampled DNN control schemes.}, archive = {J_AUTOM}, author = {Renjie Ma and Zhijian Hu and Rongni Yang and Ligang Wu}, doi = {10.1016/j.automatica.2025.112573}, journal = {Automatica}, month = {1}, pages = {112573}, shortjournal = {Automatica}, title = {Aperiodic-sampled neural network controllers with closed-loop stability verifications}, volume = {183}, year = {2026}, }
• (2026). Backstepping for partial differential equations: A survey. AUTOM, 183, 112572. (www)

  @article{ , abstract = {Systems modeled by partial differential equations (PDEs) are at least as ubiquitous as those by nature finite-dimensional and modeled by ordinary differential equations (ODEs). And yet, systematic and readily usable methodologies, for such a significant portion of real systems, have been historically scarce. Around the year 2000, the backstepping approach to PDE control began to offer not only a less abstract alternative to PDE control techniques replicating optimal and spectrum assignment techniques of the 1960s, but also enabled the methodologies of adaptive and nonlinear control, matured in the 1980s and 1990s, to be extended from ODEs to PDEs, allowing feedback synthesis for systems that are uncertain, nonlinear, and infinite-dimensional. The PDE backstepping literature has since grown to hundreds of papers and nearly a dozen books. This survey aims to facilitate the entry into this thriving area of overwhelming size and topical diversity. Designs of controllers and observers, for parabolic, hyperbolic, and other classes of PDEs, in one or more dimensions, with nonlinear, adaptive, sampled-data, and event-triggered extensions, are covered in the survey. The lifeblood of control are technology and physics. The survey places a particular emphasis on applications that have motivated the development of the theory and which have benefited from the theory and designs: flows, flexible structures, materials, thermal and chemically reacting dynamics, energy (from oil drilling to batteries and magnetic confinement fusion), and vehicles.}, archive = {J_AUTOM}, author = {Rafael Vazquez and Jean Auriol and Federico Bribiesca-Argomedo and Miroslav Krstic}, doi = {10.1016/j.automatica.2025.112572}, journal = {Automatica}, month = {1}, pages = {112572}, shortjournal = {Automatica}, title = {Backstepping for partial differential equations: A survey}, volume = {183}, year = {2026}, }
• (2026). Maximum principle for partial information non-zero sum stochastic differential games with mixed delays. AUTOM, 183, 112570. (www)

  @article{ , abstract = {This paper is concerned with one kind of partial information non-zero sum stochastic differential game with mixed delays. Both the state and control processes contain delays, where the former contains moving-average delay, discrete delay and noisy memory. We establish a necessary as well as two sufficient stochastic maximum principles for the game. As one of the main features of this research, a new kind of sufficient maximum principle is given, where the diffusion term can be controlled with non-convex control domains, and no second-order adjoint equation is needed. The theoretical results are applied to study two examples where the adjoint processes can be derived by two approaches and then the equilibrium points are obtained. This research generalizes those of stochastic optimal control problems.}, archive = {J_AUTOM}, author = {Pan Chen and Feng Zhang}, doi = {10.1016/j.automatica.2025.112570}, journal = {Automatica}, month = {1}, pages = {112570}, shortjournal = {Automatica}, title = {Maximum principle for partial information non-zero sum stochastic differential games with mixed delays}, volume = {183}, year = {2026}, }
• (2026). Finite and fixed-time feedback-based continuous-time optimization. AUTOM, 183, 112569. (www)

  @article{ , abstract = {This article presents a framework for finite and fixed-time control, using feedback-based optimization to drive a system toward its optimal operating point. Unlike traditional methods that rely on predefined set points, this approach employs endogenous control inputs derived through numerical optimization while adhering to system constraints. The controller’s dynamics are modeled as a convergent gradient flow, allowing the system to autonomously achieve its optimum without external references. The proposed control architecture guarantees finite- and fixed-time stability, with a Lyapunov-based analysis determining permissible perturbation bounds to ensure robust performance. The effectiveness of the proposed control strategy is demonstrated through simulations on a coupled-tank and a buck converter systems, successfully achieving the desired steady-state operation with the designed control approach.}, archive = {J_AUTOM}, author = {Baby Diana and Sunidhi Pandey and Shyam Kamal and Thach Ngoc Dinh}, doi = {10.1016/j.automatica.2025.112569}, journal = {Automatica}, month = {1}, pages = {112569}, shortjournal = {Automatica}, title = {Finite and fixed-time feedback-based continuous-time optimization}, volume = {183}, year = {2026}, }
• (2026). State estimation for lithium-ion batteries based on electrolyte–electrode PDE observers. AUTOM, 183, 112568. (www)

  @article{ , abstract = {Accurate information about the states of an electrochemical battery model facilitates a deeper understanding of battery behavior and enables performance enhancement. This paper first proposes backstepping Partial Differential Equation (PDE) observers for lithium concentration in the electrolyte phase within the negative electrode, positive electrode, and separator. Reverse sensitivity analysis is conducted to identify the most suitable measurable parameter for obtaining the electrolyte lithium concentration at the boundaries, which is used in the design of the electrolyte-phase observer. Subsequently, enhanced observers for the solid-phase lithium concentration in the negative and positive electrodes are developed. The proposed solid-phase observer enables more accurate State-of-Charge (SoC) estimation by leveraging the closed-loop electrolyte-phase observer. Simulations of the reverse sensitivity analysis and state observers are performed on a commercial cylindrical lithium iron phosphate ( LiFePO 4 ) cell to validate the effectiveness of the proposed approach.}, archive = {J_AUTOM}, author = {Sara Sepasiahooyi and Shu-Xia Tang}, doi = {10.1016/j.automatica.2025.112568}, journal = {Automatica}, month = {1}, pages = {112568}, shortjournal = {Automatica}, title = {State estimation for lithium-ion batteries based on electrolyte–electrode PDE observers}, volume = {183}, year = {2026}, }
• (2026). Interval-constraint multiagent systems: Global attractivity and structural stability of equilibria. AUTOM, 183, 112566. (www)

  @article{ , abstract = {This paper focuses on the dynamic behavior of an interval-constraint multiagent system. Each agent has a constraint interval that limits its potential consensus values, achieved by encoding a nonsmooth piecewise function into the agent. In addition, the underlying graphs considered are strongly connected. First, a dichotomy of equilibria is identified: either a unique non-consensus equilibrium point or multiple consensus points, depending on whether the intersection of the constraint intervals is empty or not. Then, the set of equilibria is proven to be a global attractor. Structural stability of such a system is also proven based on real-analysis methods, showing that the equilibria have continuous dependence on changes of the constraint intervals. Three running examples are used to illustrate the proposed results.}, archive = {J_AUTOM}, author = {Fengqiu Liu and Kuize Zhang and Yuhu Wu and Xiaoping Xue}, doi = {10.1016/j.automatica.2025.112566}, journal = {Automatica}, month = {1}, pages = {112566}, shortjournal = {Automatica}, title = {Interval-constraint multiagent systems: Global attractivity and structural stability of equilibria}, volume = {183}, year = {2026}, }
• (2026). Closed-loop data-enabled predictive control and its equivalence with closed-loop subspace predictive control. AUTOM, 183, 112556. (www)

  @article{ , abstract = {Factors like growing data availability and increasing system complexity have sparked interest in data-driven predictive control (DDPC) methods like Data-enabled Predictive Control (DeePC). However, closed-loop identification bias arises in the presence of noise, which reduces the effectiveness of obtained control policies. In this paper we propose Closed-loop Data-enabled Predictive Control (CL-DeePC), a framework that unifies different approaches to address this challenge. To this end, CL-DeePC incorporates instrumental variables (IVs) to synthesize and sequentially apply consistent single or multi-step-ahead predictors. Furthermore, a computationally efficient CL-DeePC implementation is developed that reveals an equivalence with Closed-loop Subspace Predictive Control (CL-SPC). Time marching simulations of DeePC and CL-DeePC are conducted using Hankel matrices of past data that are updated at every time step to induce potentially troublesome closed-loop correlations between inputs and noise. Compared to DeePC, CL-DeePC simulations demonstrate superior reference tracking, with a sensitivity study finding a 48% lower susceptibility to noise-induced reference tracking performance degradation.}, archive = {J_AUTOM}, author = {Rogier Dinkla and Tom Oomen and Sebastiaan Paul Mulders and Jan-Willem van Wingerden}, doi = {10.1016/j.automatica.2025.112556}, journal = {Automatica}, month = {1}, pages = {112556}, shortjournal = {Automatica}, title = {Closed-loop data-enabled predictive control and its equivalence with closed-loop subspace predictive control}, volume = {183}, year = {2026}, }
• (2026). Byzantine-resilient federated online learning for gaussian process regression. AUTOM, 183, 112554. (www)

  @article{ , abstract = {In this paper, we study Byzantine-resilient federated online learning for Gaussian process regression (GPR). We develop a Byzantine-resilient federated GPR algorithm that allows a cloud and a group of agents to collaboratively learn a latent function and improve the learning performances where some agents exhibit Byzantine failures, i.e., arbitrary and potentially adversarial behavior. Each agent-based local GPR sends potentially compromised local predictions to the cloud, and the cloud-based aggregated GPR computes a global model by a Byzantine-resilient product of experts aggregation rule. Then the cloud broadcasts the current global model to all the agents. Agent-based fused GPR refines local predictions by fusing the received global model with that of the agent-based local GPR. Moreover, we quantify the learning accuracy improvements of the agent-based fused GPR over the agent-based local GPR. Experiments on a toy example and two medium-scale real-world datasets are conducted to demonstrate the performances of the proposed algorithm.}, archive = {J_AUTOM}, author = {Xu Zhang and Zhenyuan Yuan and Minghui Zhu}, doi = {10.1016/j.automatica.2025.112554}, journal = {Automatica}, month = {1}, pages = {112554}, shortjournal = {Automatica}, title = {Byzantine-resilient federated online learning for gaussian process regression}, volume = {183}, year = {2026}, }
• (2026). Observer-based asymptotic active fault diagnosis: Separate and joint design of observer gain and input. AUTOM, 183, 112548. (www)

  @article{ , abstract = {This paper proposes observer gain and input design methods for observer-based asymptotic active fault diagnosis, which are based on a newly-defined notion named the excluding degree of the origin from a zonotope. Using the excluding degree, a quantitative specification is obtained to characterize the performance of set-based robust fault diagnosis. Furthermore, a separate gain design method and a joint gain and input design method are proposed, respectively. This is the first work to achieve a joint observer gain and input design for set-based active fault diagnosis. Compared with the existing methods that design gains and input separately, the proposed joint gain and input design method has advantages to exploit the fault diagnosis potential of observer-based schemes. Finally, several examples are used to illustrate the effectiveness of the proposed methods.}, archive = {J_AUTOM}, author = {Feng Xu and Yiming Wan and Ye Wang and Vicenç Puig}, doi = {10.1016/j.automatica.2025.112548}, journal = {Automatica}, month = {1}, pages = {112548}, shortjournal = {Automatica}, title = {Observer-based asymptotic active fault diagnosis: Separate and joint design of observer gain and input}, volume = {183}, year = {2026}, }
• (2026). Controller synthesis from noisy-input noisy-output data. AUTOM, 183, 112545. (www)

  @article{ , abstract = {We consider the problem of synthesizing a dynamic output-feedback controller for a linear system, using solely input–output data corrupted by measurement noise. To handle input–output data, an auxiliary representation of the original system is utilized. By exploiting the structure of the auxiliary system, we design a controller that robustly stabilizes all possible systems consistent with data. Notably, we also provide a novel solution to extend the results to generic multi-input multi-output systems. The findings are illustrated by numerical examples.}, archive = {J_AUTOM}, author = {Lidong Li and Andrea Bisoffi and Claudio De Persis and Nima Monshizadeh}, doi = {10.1016/j.automatica.2025.112545}, journal = {Automatica}, month = {1}, pages = {112545}, shortjournal = {Automatica}, title = {Controller synthesis from noisy-input noisy-output data}, volume = {183}, year = {2026}, }
• (2026). Accuracy bounds for the simulation of a class of continuous-time nonlinear models. AUTOM, 183, 112543. (www)

  @article{ , abstract = {Real-world dynamic systems evolve in the continuous-time world, while their models are simulated in the digital world using discrete-time numerical simulation algorithms. Such simulation is essential for a variety of system and control problems such as system identification and performance analysis of (control) systems. Ideally, the simulated model response should be identical to the system response. However, this is typically not the case in practice, even when the effects of unmodelled dynamics and parametric uncertainty are excluded. Even in that scenario, a mismatch exists between the response of the system and the model due to the interface between the physical world and the digital computer, unknown disturbances, and simulation inaccuracies. For the class of continuous-time, nonlinear Lur’e-type systems, this paper analyses the mismatch between the steady-state system response and the steady-state model response computed using the so-called mixed time–frequency algorithm. Firstly, a bound on the mismatch between the steady-state system response and the computed steady-state model response based on continuous-time signals is derived. Secondly, a bound for the same mismatch is derived for a sampled version of the signals. The bounds are further decomposed into several components, each given an interpretation that can be used to reduce the bounds on the mismatch. In a numerical case study, we show that reducing the bounds also reduces the actual mismatch.}, archive = {J_AUTOM}, author = {Fahim Shakib and Johan Schoukens and Alexander Yu. Pogromsky and Alexey Pavlov and Nathan van de Wouw}, doi = {10.1016/j.automatica.2025.112543}, journal = {Automatica}, month = {1}, pages = {112543}, shortjournal = {Automatica}, title = {Accuracy bounds for the simulation of a class of continuous-time nonlinear models}, volume = {183}, year = {2026}, }
• (2026). Revisiting lossless convexification: Theoretical guarantees for discrete-time optimal control problems. AUTOM, 183, 112537. (www)

  @article{ , abstract = {Lossless Convexification (LCvx) is a modeling approach that transforms a class of nonconvex optimal control problems, where nonconvexity primarily arises from control constraints, into convex problems through convex relaxations. These convex problems can be solved using polynomial-time numerical methods after discretization, which converts the original infinite-dimensional problem into a finite-dimensional one. However, existing LCvx theory is limited to continuous-time optimal control problems, as the equivalence between the relaxed convex problem and the original nonconvex problem holds only in continuous-time. This paper extends LCvx theory to discrete-time optimal control problems by classifying them into normal and long-horizon cases. For normal cases, after an arbitrarily small perturbation to the system dynamics (recursive equality constraints), applying the existing LCvx method to discrete-time problems results in optimal controls that meet the original nonconvex constraints at all but no more than n x − 1 temporal grid points, where n x is the state dimension. For long-horizon cases, the existing LCvx method fails, but we resolve this issue by integrating it with a bisection search, leveraging the continuity of the value function from the relaxed convex problem to achieve similar results as in normal cases. This paper strengthens the theoretical foundation of LCvx, extending the applicability of LCvx theory to discrete-time optimal control problems.}, archive = {J_AUTOM}, author = {Dayou Luo and Kazuya Echigo and Behçet Açıkmeşe}, doi = {10.1016/j.automatica.2025.112537}, journal = {Automatica}, month = {1}, pages = {112537}, shortjournal = {Automatica}, title = {Revisiting lossless convexification: Theoretical guarantees for discrete-time optimal control problems}, volume = {183}, year = {2026}, }
• (2026). Federated cubic regularized newton learning with sparsification-amplified differential privacy. AUTOM, 183, 112531. (www)

  @article{ , abstract = {This paper explores the cubic-regularized Newton method within a federated learning framework while addressing two major concerns: privacy leakage and communication bottlenecks. We propose the Differentially Private Federated Cubic Regularized Newton (DP-FCRN) algorithm, which leverages second-order techniques to achieve lower iteration complexity than first-order methods. We incorporate noise perturbation during local computations to ensure privacy. Furthermore, we employ sparsification in uplink transmission, which not only reduces the communication costs but also amplifies the privacy guarantee. Specifically, this approach reduces the necessary noise intensity without compromising privacy protection. We analyze the convergence properties of our algorithm and establish the privacy guarantee. Finally, we validate the effectiveness of the proposed algorithm through experiments on a benchmark dataset.}, archive = {J_AUTOM}, author = {Wei Huo and Changxin Liu and Kemi Ding and Karl Henrik Johansson and Ling Shi}, doi = {10.1016/j.automatica.2025.112531}, journal = {Automatica}, month = {1}, pages = {112531}, shortjournal = {Automatica}, title = {Federated cubic regularized newton learning with sparsification-amplified differential privacy}, volume = {183}, year = {2026}, }
• (2026). A consensus kalman filter on l2 spaces. AUTOM, 183, 112530. (www)

  @article{ , abstract = {We study the estimation problem of infinite dimensional discrete-time stochastic linear systems with finite dimensional measurements on sensor networks modeled by connected undirected graphs. The framework encompasses discretized PDEs with sampled measurements. A new scheme of distributed consensus on measurements is extended to systems evolving in L 2 spaces in order to limit the information exchange to finite-dimensional vectors. We show that, in analogy to the finite-dimensional case, at each node the variance of the estimation error tends to the one of the centralized Kalman filter for systems is L 2 when the number of consensus steps increases.}, archive = {J_AUTOM}, author = {Stefano Battilotti and Alessandro Borri and Filippo Cacace and Massimiliano d’Angelo and Alfredo Germani}, doi = {10.1016/j.automatica.2025.112530}, journal = {Automatica}, month = {1}, pages = {112530}, shortjournal = {Automatica}, title = {A consensus kalman filter on l2 spaces}, volume = {183}, year = {2026}, }
• (2026). Fixed-time learning for safe time-critical verification using reachability analysis. AUTOM, 183, 112528. (www)

  @article{ , abstract = {In this paper, we address a safe time-critical control problem using reachability analysis and design a reinforcement learning-based mechanism for learning online and in fixed-time the solution to the safe time-critical control problem. Safety is ensured by determining a set of states for which there exists an admissible control law generating a system trajectory that does not reach a set of forbidden states at a user-prescribed time instant. Specifically, we cast our safe time-critical problem as a Mayer optimal feedback control problem whose solution satisfies the Hamilton–Jacobi–Bellman (HJB) equation and characterizes the set of safe states. Since the HJB equation is generally difficult to solve, we develop an online critic-only reinforcement learning-based algorithm for simultaneously learning the solution to the HJB equation and the safe set in a fixed time. In particular, we introduce a non-Lipschitz experience replay-based learning law utilizing recorded and current data for updating the critic weights to learn the value function and the safe set. The non-Lipschitz property of the dynamics gives rise to fixed-time convergence, whereas the experience replay-based approach eliminates the need to satisfy the persistence of excitation condition provided that a recorded data set is sufficiently rich. Simulation results illustrate the efficacy of the proposed approach to the problem of fixed-wing unmanned aerial vehicle collision avoidance.}, archive = {J_AUTOM}, author = {Nick-Marios T. Kokolakis and Kyriakos G. Vamvoudakis and Wassim M. Haddad}, doi = {10.1016/j.automatica.2025.112528}, journal = {Automatica}, month = {1}, pages = {112528}, shortjournal = {Automatica}, title = {Fixed-time learning for safe time-critical verification using reachability analysis}, volume = {183}, year = {2026}, }
• (2026). Orchestrating on-board sensors for global hybrid robust stabilization of unicycles. AUTOM, 183, 112502. (www)

  @article{ , abstract = {We consider mobile robots described through unicycle dynamics equipped with on-board range sensors and cameras, one facing forward and one facing backward, providing measurements of the distance and misalignment to a target. We propose a hybrid control law combining the two on-board measurements and discuss stability results for the closed-loop expressed in the on-board camera-based coordinates, using Lyapunov-based arguments. We prove robustness of the stability properties to uncertainties affecting the sensors and external perturbations acting on the robot. The results are illustrated via simulations.}, archive = {J_AUTOM}, author = {Riccardo Ballaben and Alessandro Astolfi and Philipp Braun and Luca Zaccarian}, doi = {10.1016/j.automatica.2025.112502}, journal = {Automatica}, month = {1}, pages = {112502}, shortjournal = {Automatica}, title = {Orchestrating on-board sensors for global hybrid robust stabilization of unicycles}, volume = {183}, year = {2026}, }

CMA - 5

• (2026). Numerical simulation of generalized finite difference wave equation based on automatic construction of interface nodes. CMA, 201, 65-79. (www)

  @article{ , abstract = {In seismic exploration, the Finite Difference Method (FDM) is commonly used to simulate the propagation of seismic waves. This numerical method approximates derivatives by calculating the weighted sum of values at adjacent nodes. Currently, models are stored using regular grids. When discretizing complex structural interfaces such as inclined strata and faults, non - smooth surfaces will appear, which leads to errors in the forward modeling of wave equations. To address the forward - modeling issues caused by model errors, a meshless method called the Generalized Finite Difference Method (GFDM) has been proposed and has been successfully applied in computational geosciences. This method constructs node clouds for discrete nodes and solves derivatives based on the relationships between central nodes and surrounding nodes in the node cloud, offering flexible solutions. However, it requires substantial time for fine-grained modeling of the model, and conventional node construction performs poorly for complex models with drastic lateral variations. In this study, we analyzed the causes of errors in relation to interface dip angles and proposed an automatic interface node construction method. Testing on various models demonstrates that this method can effectively construct models with smaller discretization errors and significantly reduced scattering errors in forward modeling compared to conventional methods.}, archive = {J_CMA}, author = {Zhili Chen and Yong Wang and Zhixian Gui}, doi = {10.1016/j.camwa.2025.10.001}, journal = {Computers & Mathematics with Applications}, month = {1}, pages = {65-79}, shortjournal = {Comput. Meth. Appl.}, title = {Numerical simulation of generalized finite difference wave equation based on automatic construction of interface nodes}, volume = {201}, year = {2026}, }
• (2026). Meshfree method for solving the elliptic monge-ampère equation with dirichlet boundary. CMA, 201, 53-64. (www)

  @article{ , abstract = {We develop the meshfree method for solving the elliptic Monge-Ampère equation with Dirichlet boundary on the bounded domain and prove its convergence in the paper. In terms of trial, we use the radial functions (for example, Whittle-Matérn-Sobolev kernels and Wendland's compactly supported radial basis functions) which can reproduce W σ , 2 ( R d ) to construct the finite dimensional approximate spaces. It allows the easy construction of approximation spaces in arbitrary dimensions with arbitrary smoothness and avoids the huge workload caused by mesh-based methods at the same time. In terms of testing, it just needs to take values directly on the collocation points and greatly simplifies the difficulties caused by variation and integration. Theoretically, we obtain L 2 error when the testing discretization is finer than the trial discretization. The convergence rate depends on the regularity of the solution, the smoothness of the computing domain, and the approximation of kernel-based trial spaces. An extension to non-Dirichlet boundary condition is in a forthcoming paper.}, archive = {J_CMA}, author = {Zhiyong Liu and Qiuyan Xu}, doi = {10.1016/j.camwa.2025.09.036}, journal = {Computers & Mathematics with Applications}, month = {1}, pages = {53-64}, shortjournal = {Comput. Meth. Appl.}, title = {Meshfree method for solving the elliptic monge-ampère equation with dirichlet boundary}, volume = {201}, year = {2026}, }
• (2026). A second-order time-accurate, linear fully decoupled unconditional energy stabilization finite element method for tumor growth model. CMA, 201, 35-52. (www)

  @article{ , abstract = {By using the L 2 gradient flow method, we derive a phase-field model for tumor growth from the free energy. The scalar auxiliary variable (SAV) method is employed to handle the nonlinear energy potential. Based on the second-order backward differentiation formula (BDF2) and the finite element method, we construct an unconditionally stable, linear, and decoupled fully discrete numerical scheme. We rigorously prove the unconditional energy stability of the proposed scheme and the optimal L 2 -norm error estimates for ϕ and c . Numerical examples are presented to validate the theoretical results and to demonstrate the effectiveness of the model and the scheme.}, archive = {J_CMA}, author = {Mingle Sun and Bo Wang and Guang-an Zou and Yuxing Zhang}, doi = {10.1016/j.camwa.2025.09.028}, journal = {Computers & Mathematics with Applications}, month = {1}, pages = {35-52}, shortjournal = {Comput. Meth. Appl.}, title = {A second-order time-accurate, linear fully decoupled unconditional energy stabilization finite element method for tumor growth model}, volume = {201}, year = {2026}, }
• (2026). High order energy invariant fast algorithm for space two dimensional klein-gordon-zakharov equations. CMA, 201, 18-34. (www)

  @article{ , abstract = {Space two dimensional Klein-Gordon-Zakharov equations are directly changed into the Hamiltonian system with infinite dimensional space by the variational formula, which can be discretized into finite dimensional Hamiltonian system by Fourier pseudo-spectral method. The average vector field formulas with second and fourth order accuracy in time are utilized to compute the finite dimensional Hamiltonian system. In order to improve computation velocity of these formulas, the fast computation algorithm of these formulas is proposed by decomposing the spectral matrix. Solitary wave evolution of the equations is analyzed with different initial conditions by these new computational formulas. Energy invariant property, accuracy and computation efficiently of these new formulas are also investigated.}, archive = {J_CMA}, author = {Jie Chen and Jianqiang Sun}, doi = {10.1016/j.camwa.2025.09.026}, journal = {Computers & Mathematics with Applications}, month = {1}, pages = {18-34}, shortjournal = {Comput. Meth. Appl.}, title = {High order energy invariant fast algorithm for space two dimensional klein-gordon-zakharov equations}, volume = {201}, year = {2026}, }
• (2026). Fundamental solution neural networks for solving inverse cauchy problems for the laplace and biharmonic equations. CMA, 201, 1-17. (www)

  @article{ , abstract = {This paper proposes a novel fundamental solution neural networks method (FSNNs) to solve inverse Cauchy problems, which combines the method of fundamental solutions (MFS) with the physics-informed neural networks (PINNs). To optimize the distribution of source points, the method starts by partitioning the interval into equal segments, determined by the number of the source points. The coordinate system is situated at the center of the computational domain. The resulting angles as network inputs for the FSNNs and the intermediate variable as outputs, which is subsequently substituted into a length function to obtain the final length. The coordinates of the source points are then determined, and the MFS is employed to approximate the numerical solutions. The loss function is formulated based on the boundary conditions on the accessible boundary, and the training is employed to optimize the network parameters in the FSNNs and source point intensities in the MFS. The introduction of the PINNs overcomes the challenge of source point selection in the MFS and effectively addresses the ill-posedness of inverse problems. In summary, the proposed scheme is a machine learning-based semi-analytical meshless method which is simple, accurate and easily implemented, making it highly suitable for the numerical solution of inverse problems. Four numerical experiments, including the Laplace and biharmonic equations, validate the effectiveness and accuracy of the proposed FSNNs.}, archive = {J_CMA}, author = {Xin Li and Fajie Wang and Renhao Wang and Shengdong Zhao and Daigeng Yang}, doi = {10.1016/j.camwa.2025.09.032}, journal = {Computers & Mathematics with Applications}, month = {1}, pages = {1-17}, shortjournal = {Comput. Meth. Appl.}, title = {Fundamental solution neural networks for solving inverse cauchy problems for the laplace and biharmonic equations}, volume = {201}, year = {2026}, }

CMAME - 39

• (2026). Non-iterative method for reliability-based topology optimization of static and dynamic problems with uncertain-but-bounded parameters. CMAME, 448, 118456. (www)

  @article{ , abstract = {The non-probabilistic reliability-based topology optimization (NRBTO) method offers a powerful tool for achieving high-performance and high-safety layout designs under uncertain-but-bounded (UBB) parameters. However, the practical application of NRBTO is limited by unaffordable computational burden associated with repeated non-probabilistic reliability iterations. To address this issue, a new non-iterative reliability-based topology optimization method for static and dynamic problems with UBB parameters is established to avoid the non-probabilistic reliability iteration, in which a novel non-iterative method (NIM) is proposed by strict proving the monotonicity conditions with respect to UBB parameters. The complex NRBTO model is transformed into an equivalent deterministic topology optimization model, thereby reducing computational costs. Moreover, the sensitivities with respect to design variables are derived using the adjoint method. The effectiveness of proposed NIM for NRBTO is validated through two static and two dynamic examples, where the compliance, displacement, stress, frequency, and frequency-band problems are addressed.}, archive = {J_CMAME}, author = {Gang Yang and Zeng Meng and Peng Hao and Changting Zhong}, doi = {10.1016/j.cma.2025.118456}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118456}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Non-iterative method for reliability-based topology optimization of static and dynamic problems with uncertain-but-bounded parameters}, volume = {448}, year = {2026}, }
• (2026). Aerodynamics-driven topology optimization of compliant airfoils considering stability. CMAME, 448, 118454. (www)

  @article{ , abstract = {Airfoil structures optimized solely for stiffness can suffer from buckling instabilities under realistic aerodynamic loads. We present the first topology optimization framework to improve the stability of aerodynamic structures. For a clear representation of structure, this work employs the topology optimization of binary structures with geometry trimming. Reynolds-averaged Navier-Stokes turbulence model is employed to accurately predict the turbulent aerodynamic loading under realistic flight conditions. Fluid–structure interaction and buckling analysis are conducted using an elastic formulation with geometrical nonlinearities to allow for large deformations. The numerical model system is solved through the finite element method and the Arbitrary Lagrangian-Eulerian method is applied. The sensitivities are calculated using semi-automatic differentiation and interpolated to the optimization mesh. Kreisselmeier-Steinhauser aggregation function is used and augmented Lagrangian multipliers are developed for buckling constraints. Numerical examples demonstrate that the proposed method can effectively improve the airfoil stability to different constraint levels across various configurations with minimal parameter tuning. Additionally, the algorithm produces designs that are conducive to manufacturing.}, archive = {J_CMAME}, author = {Qingdi Wang and Lucas Oliveira Siqueira and Tao Xu and Guanzhe Cui and Zhi Li and Anderson Soares da Costa Azevêdo and Renato Picelli and Yi Min Xie}, doi = {10.1016/j.cma.2025.118454}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118454}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Aerodynamics-driven topology optimization of compliant airfoils considering stability}, volume = {448}, year = {2026}, }
• (2026). Nine circles of elastic brittle fracture: A series of challenge problems to assess fracture models. CMAME, 448, 118449. (www)

  @article{ , abstract = {Since the turn of the millennium, capitalizing on modern advances in mathematics and computation, a slew of computational models have been proposed in the literature with the objective of describing the nucleation and propagation of fracture in materials subjected to mechanical, thermal, and/or other types of loads. By and large, each new proposal focuses on a particular aspect of the problem, while ignoring others that have been well-established. This approach has resulted in a plethora of models that are, at best, descriptors of fracture only under a restricted set of conditions, while they may predict grossly incorrect and even non-physical behaviors in general. In an attempt to address this predicament, this paper introduces a vetting process in the form of nine challenge problems that any computational model of fracture must convincingly handle if it is to potentially describe fracture nucleation and propagation in general. The focus is on the most basic of settings, that of isotropic elastic brittle materials subjected to quasi-static mechanical loads. The challenge problems have been carefully selected so that: i ) they can be carried out experimentally with standard testing equipment; i i ) they can be unambiguously analyzed with a sharp description of fracture; and, most critically, i i i ) in aggregate they span the entire range of well settled experimental knowledge on fracture nucleation and propagation that has been amassed for over a century. For demonstration purposes, after their introduction, each challenge problem is solved with two phase-field models of fracture, a classical variational phase-field model and the phase-field model initiated by Kumar, Francfort, and Lopez-Pamies ( J. Mech. Phys. Solids 112 (2018), 523–551), this both for a prototypical elastic brittle hard material (soda-lime glass) and a prototypical elastic brittle soft material (a polyurethane elastomer).}, archive = {J_CMAME}, author = {Farhad Kamarei and Bo Zeng and John E. Dolbow and Oscar Lopez-Pamies}, doi = {10.1016/j.cma.2025.118449}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118449}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Nine circles of elastic brittle fracture: A series of challenge problems to assess fracture models}, volume = {448}, year = {2026}, }
• (2026). Nonlinear projection-based model order reduction with machine learning regression for closure error modeling in the latent space. CMAME, 448, 118443. (www)

  @article{ , abstract = {A significant advancement in nonlinear projection-based model order reduction (PMOR) is presented through a highly effective methodology. This methodology employs Gaussian process regression (GPR) and radial basis function (RBF) interpolation for closure error modeling in the latent space, offering notable gains in efficiency and expanding the scope of PMOR. Moving beyond the limitations of deep artificial neural networks (ANNs), previously used for this task, this approach provides crucial advantages in terms of interpretability and a reduced demand for extensive training data. The capabilities of GPR and RBFs are showcased in two demanding applications: a two-dimensional parametric inviscid Burgers problem, featuring propagating shocks across the entire computational domain, and a complex three-dimensional turbulent flow simulation around an Ahmed body. The results demonstrate that this innovative approach preserves accuracy and achieves substantial improvements in efficiency and interpretability when contrasted with traditional PMOR and ANN-based closure modeling.}, archive = {J_CMAME}, author = {S. Ares de Parga and Radek Tezaur and Carlos G. Hernández and Charbel Farhat}, doi = {10.1016/j.cma.2025.118443}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118443}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Nonlinear projection-based model order reduction with machine learning regression for closure error modeling in the latent space}, volume = {448}, year = {2026}, }
• (2026). A block preconditioner for thermo-poromechanics with frictional deformation of fractures. CMAME, 448, 118440. (www)

  @article{ , abstract = {The numerical modeling of fracture contact thermo-poromechanics is crucial for advancing subsurface engineering applications, including CO 2 sequestration, production of geo-energy resources, energy storage and wastewater disposal operations. Accurately modeling this problem presents substantial challenges due to the complex physics involved in strongly coupled thermo-poromechanical processes and the frictional contact mechanics of fractures. To resolve process couplings in the resulting mathematical model, it is common to apply fully implicit time stepping. This necessitates the use of an iterative linear solver to run the model. The solver’s efficiency primarily depends on a robust preconditioner, which is particularly challenging to develop because it must handle the mutual couplings between linearized contact mechanics and energy, momentum, and mass balance. In this work, we introduce a preconditioner for the problem based on the nested approximations of Schur complements. To decouple the momentum balance, we utilize the fixed-stress approximation, extended to account for both the porous media and fracture subdomains. The singularity of the contact mechanics submatrix is resolved by a linear transformation. Two variations of the algorithm are proposed to address the coupled mass and energy balance submatrix: either the Constrained Pressure Residual (CPR) or the algebraic multigrid method (AMG) for systems of equations. The preconditioner is evaluated through numerical experiments of fluid injection into fractured porous media, which causes thermal contraction and subsequent sliding and opening of fractures. The experiments show that the preconditioner performs robustly for a wide range of simulation regimes governed by various fracture states, friction coefficients and Peclétnumber. The grid refinement experiments demonstrate that the preconditioner scales well in terms of GMRES iterations, in both two and three dimensions.}, archive = {J_CMAME}, author = {Yury Zabegaev and Inga Berre and Eirik Keilegavlen}, doi = {10.1016/j.cma.2025.118440}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118440}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {A block preconditioner for thermo-poromechanics with frictional deformation of fractures}, volume = {448}, year = {2026}, }
• (2026). Inverse scattering without phase: Carleman convexification and phase retrieval via the Wentzel–Kramers–Brillouin approximation. CMAME, 448, 118439. (www)

  @article{ , abstract = {This paper addresses the challenging and interesting inverse problem of reconstructing the spatially varying dielectric constant of a medium from phaseless backscattering measurements generated by single-point illumination. The underlying mathematical model is governed by the three-dimensional Helmholtz equation, and the available data consist solely of the magnitude of the scattered wave field. To address the nonlinearity and servere ill-posedness of this phaseless inverse scattering problem, we introduce a robust, globally convergent numerical framework combining several key regularization strategies. Our method first employs a phase retrieval step based on the Wentzel–Kramers–Brillouin (WKB) ansatz, where the lost phase information is reconstructed by solving a nonlinear optimization problem. Subsequently, we implement a Fourier-based dimension reduction technique, transforming the original problem into a more stable system of elliptic equations with Cauchy boundary conditions. To solve this resulting system reliably, we apply the Carleman convexification approach, constructing a strictly convex weighted cost functional whose global minimizer provides an accurate approximation of the true solution. Numerical simulations using synthetic data with high noise levels demonstrate the effectiveness and robustness of the proposed method, confirming its capability to accurately recover both the geometric location and contrast of hidden scatterers.}, archive = {J_CMAME}, author = {Thuy T. Le and Phuong M. Nguyen and Loc H. Nguyen}, doi = {10.1016/j.cma.2025.118439}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118439}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Inverse scattering without phase: Carleman convexification and phase retrieval via the Wentzel–Kramers–Brillouin approximation}, volume = {448}, year = {2026}, }
• (2026). A green’s function fast multipole method for computation of micromechanical fields in heterogeneous materials. CMAME, 448, 118436. (www)

  @article{ , abstract = {Computation of micromechanical fields in heterogeneous materials is usually performed using either the finite element method or the Green’s function method based on FFTs. The finite element method allows for accurate discretization and for non-periodic boundary conditions but is computationally expensive. On the other hand, the FFT-based method is computationally efficient but requires discretization on a regular grid of hexahedral voxels. In this paper, a Green’s function method allowing for accurate discretization using tetrahedral elements and for non-periodic boundary conditions is proposed. The convolution is computed using the fast multipole method, which provides good accuracy even for low-order expansion due to the fast decay of interactions between elements. The proposed Green’s function fast multipole method is verified by comparison with analytical and FFT-based solutions. The computational time is analyzed and compared to the FFT-based method for non-periodic convolution. Finally, effective properties of an elastic polycrystalline microstructure containing thin intergranular cracks are computed and analyzed.}, archive = {J_CMAME}, author = {Miroslav Zecevic}, doi = {10.1016/j.cma.2025.118436}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118436}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {A green’s function fast multipole method for computation of micromechanical fields in heterogeneous materials}, volume = {448}, year = {2026}, }
• (2026). A mesh-based geometric deep learning framework for rapid response prediction of large-scale and multi-component mechanical structures in engineering. CMAME, 448, 118435. (www)

  @article{ , abstract = {Mesh-based finite element method (FEM) plays a critical role in simulating structural response. However, the analysis of complex physical processes, such as vehicle crashworthiness, is hindered by inherent high nonlinearities and large-scale meshes, leading to significant computational overhead and impeding rapid structure design. In recent years, the use of machine learning (ML) or deep learning (DL) methods to build surrogate models for simulations has gained much attention, which offers the potential to drastically reduce computational time while preserving accuracy. In this paper, we develop an end-to-end and mesh-based geometric DL framework that takes finite element (FE) solver files (such as.k file for LS-Dyna) containing mesh and material information as input, and quickly outputs response prediction of large-scale and multi-component mechanical structures in engineering, thus serving a promising alternative to FE solvers. We innovatively introduce the graph self-supervised learning (SSL) to transform FE data of structural component with varied material properties, complex geometric shapes and arbitrary number of unstructured meshes into low-dimensional embeddings, which are then employed to build an equivalent small-scale graph representation of the large-scale assembly, effectively alleviating the computational costs of subsequent prediction models. Then, we present GNN-FNN and GNN-Transformer models specifically designed for three different prediction tasks, including forecasting static and dynamic structural performance metrics, and constructing time-dependent physical fields. Using a large-scale industrial case of the electric vehicle (EV) under side pole impact, three regression tasks are carried out to assess the effectiveness of the proposed approach. Results reveal that the non-parametric model, free from the need for manually defined explicit parameters, excels in extracting implicit parameters for diverse structures, which support satisfactory prediction accuracy in each task with a considerable speedup than the simulation. Besides, it is surprising that our model is weakly sensitive to the moderate variation in the mesh resolution, which is valuable for practical engineering applications. The adaptability and scalability of our method are further verified on three additional industrial cases with varied structural simulation scenarios and progressively increasing FE model complexity. This work offers an effective surrogate model to accelerate the response evaluation of mechanical structures in engineering and shorten the design cycle requiring iterative optimization.}, archive = {J_CMAME}, author = {Gongxi Zhang and Ying Liu and Yi Quan and Junfei Yan}, doi = {10.1016/j.cma.2025.118435}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118435}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {A mesh-based geometric deep learning framework for rapid response prediction of large-scale and multi-component mechanical structures in engineering}, volume = {448}, year = {2026}, }
• (2026). Physics- and data-driven active learning of neural network representations for free energy density functions of materials from statistical mechanics. CMAME, 448, 118434. (www)

  @article{ , abstract = {Accurate free energy density representations are crucial for understanding phase dynamics in materials. We employ a scale-bridging approach to incorporate atomistic information into our free energy density model by training a neural network on DFT-informed Monte Carlo data. To optimize sampling in the high-dimensional Monte Carlo space, we present an active learning framework that integrates space-filling sampling, uncertainty-based sampling, and physics-informed sampling. Additionally, our approach includes methods such as hyperparameter tuning, dynamic sampling, and novelty enforcement. These strategies can be combined to reduce the mean squared error-either globally or in targeted regions of interest-while minimizing the number of required data points. The framework introduced here is broadly applicable to Monte Carlo sampling of a range of materials systems.}, archive = {J_CMAME}, author = {J. Holber and K. Garikipati}, doi = {10.1016/j.cma.2025.118434}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118434}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Physics- and data-driven active learning of neural network representations for free energy density functions of materials from statistical mechanics}, volume = {448}, year = {2026}, }
• (2026). Dimension-reduced chapman-kolmogorov equation for high-dimensional stochastic dynamical systems. CMAME, 448, 118433. (www)

  @article{ , abstract = {Random vibration analysis of high-dimensional dynamical systems is a fundamental problem in science and engineering, yet it remains challenging due to the curse of dimensionality. While dimension-reduced formulations have been developed for differential-type equations governing time-variant probability density, such as the Fokker-Planck equation, no equivalent formulation has been established for the integral-type Chapman-Kolmogorov (CK) equation, despite its theoretical importance and computational advantages. In this paper, a novel dimension-reduced Chapman-Kolmogorov (DRCK) equation is established governing the transient probability density function (PDF) of any quantity of interest in high-dimensional Markov systems. The derivation is conducted based on the projection of the full Chapman-Kolmogorov equation onto the dimension-reduced space. It is established that the intrinsic transition probability density (TPD) of the DRCK equation is the conditional expectation of the original TPD. Further, the short-time approximate intrinsic TPDs under both Gaussian and Poisson white noise excitations are derived analytically, enabling practical numerical implementation. The proposed DRCK equation provides a mathematically rigorous and computationally efficient framework for high-dimensional stochastic systems. Numerical examples are developed to demonstrate its accuracy and effectiveness. The DRCK equation thus provides a new tool for reliability assessment and uncertainty quantification in complex engineering applications.}, archive = {J_CMAME}, author = {Jianbing Chen and Meng-Ze Lyu and Shenghan Zhang}, doi = {10.1016/j.cma.2025.118433}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118433}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Dimension-reduced chapman-kolmogorov equation for high-dimensional stochastic dynamical systems}, volume = {448}, year = {2026}, }
• (2026). Second order three-dimensional serendipity virtual elements for hyperelasticity: Static and dynamic analysis. CMAME, 448, 118432. (www)

  @article{ , abstract = {In this work, a three-dimensional (3D) second-order serendipity virtual element method (S-VEM) is developed for the static and dynamic analysis of hyperelastic materials. The VEM framework is based on the projection of unknown basis functions onto polynomial spaces, allowing for flexible discretization with arbitrary polyhedral meshes. While most existing VEM formulations for 3D mechanical problems are discretized using first-order formulations, higher-order schemes offer improved precision, especially for nonlinear problems. However, conventional second-order VEM formulations introduce additional degrees of freedom (DOFs), such as body and surface moments, which complicate the implementation and reduce computation efficiency. To address this challenge, we propose a novel 3D second-order serendipity VEM that avoids any extra moment-related DOFs. This is the first application of a serendipity VEM to 3D static and dynamic problems in hyperelasticity. Furthermore, by integrating advanced mesh generation techniques, the proposed method enables hybrid simulations that combine second-order serendipity VEM and FEM to efficiently handle complex geometries.}, archive = {J_CMAME}, author = {Bing-Bing Xu and Lourenco Beirao da Veiga and Yongjie Jessica Zhang and Peter Wriggers}, doi = {10.1016/j.cma.2025.118432}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118432}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Second order three-dimensional serendipity virtual elements for hyperelasticity: Static and dynamic analysis}, volume = {448}, year = {2026}, }
• (2026). Stabilized explicit material point method for fluid flow and fluid-structure interaction simulations using dual high-order B-spline volume averaging. CMAME, 448, 118428. (www)

  @article{ , abstract = {Traditional explicit Material Point Methods (MPM) for weakly compressible fluids often suffer from volumetric locking, cell-crossing instability, and excessive energy dissipation, particularly in fluid-structure interaction (FSI) scenarios. This study presents a stabilized explicit MPM framework that integrates dual high-order B-spline volume averaging to address these challenges. The proposed dual averaging technique simultaneously smooths deformation gradients and pressure fields using cubic B-spline basis functions to eliminate cell-crossing errors and reduce volumetric locking. A blended APIC/FLIP mapping scheme is developed to enhance energy conservation and stability at coarse grid resolutions. The framework is further enhanced by seamlessly integrating various complementary techniques such as δ -correction, pressure smoothing, and specialized boundary handling for more robust and effective modeling of free-surface and FSI problems. The framework is rigorously validated through benchmark cases, including 1D elastic wave propagation, Poiseuille flow, lid-driven cavity flow, water sloshing, dam break, and water impact on elastic obstacles. The simulation results demonstrate a remarkable reduction in pressure oscillations and improved particle distribution uniformity compared to prior MPM approaches. The proposed method establishes a robust and efficient tool for large-deformation FSI problems and bridges gaps in accuracy and stability for industrial-scale applications.}, archive = {J_CMAME}, author = {Zhang Cheng and Shiwei Zhao and Hao Chen and Jidong Zhao}, doi = {10.1016/j.cma.2025.118428}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118428}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Stabilized explicit material point method for fluid flow and fluid-structure interaction simulations using dual high-order B-spline volume averaging}, volume = {448}, year = {2026}, }
• (2026). Personalized multiscale modeling of coronary plaque progression: The interaction between low-density-lipoprotein transport and cellular dynamics. CMAME, 448, 118427. (www)

  @article{ , abstract = {Multiscale agent-based modeling has shown promise in elucidating the mechanobiological mechanisms underlying atherosclerotic plaque formation and progression. However, the integration of advanced models of low-density lipoprotein (LDL) transport in the lumen and across the endothelium with agent-based models (ABMs) of plaque growth remains underexplored. Furthermore, patient-specific applications are lacking. This study introduces a novel agent-based modeling framework for atherosclerosis, integrating hemodynamics and LDL transport in the lumen through computational fluid dynamics simulations, a three-pore model of trans-endothelial LDL migration, and an ABM of lipid and cellular dynamics. For the first time, the framework was applied to a patient-specific coronary artery and validated against 1-year follow-up data. Furthermore, it was used to explore potential plaque evolution over 5 years and under elevated LDL concentration. The calibrated model predicted the 1-year variation in wall area in two patient-specific coronary cross-sections with an error of less than 10%. Simulated scenarios indicated that variations in blood LDL concentrations can result in distinct plaque morphologies, from localized to diffuse patterns. This study provided an innovative, advanced multiscale model of atherosclerotic plaque formation and progression. As the first patient-specific application of a multiscale agent-based modeling framework for atherosclerosis with initial validation, this study underscored the potential of the approach for deciphering the mechanobiological pathways driving coronary plaque progression. The developed model provided valuable insights into how the interplay between LDL transport and hemodynamics influences arterial wall cellular dynamics in a patient-specific context.}, archive = {J_CMAME}, author = {Anna Corti and Giuseppe De Nisco and Jolanda J. Wentzel and Francesco Migliavacca and Umberto Morbiducci and Claudio Chiastra}, doi = {10.1016/j.cma.2025.118427}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118427}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Personalized multiscale modeling of coronary plaque progression: The interaction between low-density-lipoprotein transport and cellular dynamics}, volume = {448}, year = {2026}, }
• (2026). Linearized localized orthogonal decomposition for quasilinear nonmonotone elliptic PDE. CMAME, 448, 118426. (www)

  @article{ , abstract = {In this paper, we propose and analyze a multiscale method for a class of quasilinear elliptic problems of nonmonotone type with spatially multiscale coefficient. The numerical approach is inspired by the Localized Orthogonal Decomposition (LOD), so that we do not require structural assumptions such as periodicity or scale separation and only need minimal regularity assumptions on the coefficient. To construct the multiscale space, we solve linear fine-scale problems on small local subdomains, for which we consider two different linearization techniques. For both, we present a rigorous well-posedness analysis and convergence estimates in the H 1 -semi norm. We compare and discuss theoretically and numerically the performance of our strategies for different linearization points. Both numerical experiments and theoretical analysis demonstrate the validity and applicability of the method.}, archive = {J_CMAME}, author = {Maher Khrais and Barbara Verfürth}, doi = {10.1016/j.cma.2025.118426}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118426}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Linearized localized orthogonal decomposition for quasilinear nonmonotone elliptic PDE}, volume = {448}, year = {2026}, }
• (2026). The isogeometric MITC shell in geometric nonlinear analysis. CMAME, 448, 118425. (www)

  @article{ , abstract = {This paper extends the isogeometric MITC shell formulation proposed by Mi and Yu (2021) for linear analysis to address geometric nonlinear shell problems. Built on the Reissner-Mindlin shell theory, the original linear formulation employs the Mixed Interpolation of Tensorial Components (MITC) technique to alleviate shear and membrane locking. The present nonlinear extension retains the MITC framework while incorporating a mixed Non-Uniform Rational B-Spline (NURBS)-Lagrange interpolation strategy to address the additional complexities induced by geometric nonlinearity. The interpolatory nature of the Lagrange basis functions is leveraged to simplify the construction of director vectors and assumed strain fields. The nonlinear problem is formulated in a total Lagrangian setting and solved using Newton-Raphson iterations. The effectiveness of the proposed method is demonstrated through a comprehensive set of numerical examples, including both standard benchmarks and a collection of geometric nonlinear shell problems, which features challenging behaviors such as large rotations and the development of local creases. Through Bézier extraction, the method is further evaluated using T-spline and U-spline basis functions. The numerical results confirm that the MITC technique effectively suppresses shear and membrane locking, and the proposed shell formulation exhibits high accuracy and robust convergence, even for coarse and severely distorted meshes. However, it is also observed that the high inter-element continuity inherent in splines-based discretization can inhibit large deformations, introducing a new form of locking. This issue is successfully mitigated using Bézier extraction to reduce the inter-element continuity. Overall, the proposed formulation offers a general and reliable isogeometric framework for geometrically nonlinear analysis, applicable across a wide range of shell geometries, loading conditions, and boundary constraints.}, archive = {J_CMAME}, author = {Yongzhen Mi}, doi = {10.1016/j.cma.2025.118425}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118425}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {The isogeometric MITC shell in geometric nonlinear analysis}, volume = {448}, year = {2026}, }
• (2026). A new pre-stressing algorithm for patient-specific simulations of growth and remodeling using the homogenized constrained mixture theory. CMAME, 448, 118424. (www)

  @article{ , abstract = {The homogenized constrained mixture theory has been implemented in different Finite-Element software packages for almost ten years to simulate growth and remodeling in soft biological tissues. In these models, it is essential to determine the pre-stresses of each constituent of the mixture in the reference configuration. However, no efficient numerical solution has been proposed so far to solve this problem. We propose to address this lack with a new algorithm based on the concept of anisotropic thermal contraction. In this algorithm, the pre-stretch tensor is incrementally updated by applying a series of anisotropic thermal contractions to each representative volume element of the model to restore its reference configuration. These contractions are proportional to the inverse of the local right stretch tensor, obtained through the polar decomposition of the deformation gradient. We implemented the algorithm in the Abaqus Finite-Element package through a UMAT subroutine and verified it on examples including growth and remodeling of a patient-specific aorta. We demonstrate that the model captures the residual stresses in good agreement with experimental results. To highlight the potential clinical relevance of the proposed algorithm, we expanded our model predictions to investigate the influence of the aortic axial pre-stretch on morphological and microstructural evolutions of the aorta under hypertensive conditions. Our simulations show that loss of axial tension, induced by hypertension, increases aortic length and may lead to pathological deformations such as aortic tortuosity. These findings highlight the importance of efficiently determining pre-stresses for simulating long-term vascular growth and remodeling.}, archive = {J_CMAME}, author = {Ali Akbar Karkhaneh Yousefi and Stéphane Avril}, doi = {10.1016/j.cma.2025.118424}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118424}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {A new pre-stressing algorithm for patient-specific simulations of growth and remodeling using the homogenized constrained mixture theory}, volume = {448}, year = {2026}, }
• (2026). GPU-accelerated multi-phase, multi-resolution SPH method with ray tracing for laser powder bed fusion. CMAME, 448, 118423. (www)

  @article{ , abstract = {Powder-scale simulation of laser powder bed fusion (LPBF) is increasingly vital for understanding, predicting, and controlling metallurgical defects. However, the complex multi-physics and multi-material interactions involved, along with high computational demands, pose significant challenges. This study presents the first multiphase smoothed particle hydrodynamics (SPH) simulation framework for LPBF under both low and high evaporation regimes, incorporating a multi-resolution particle strategy and ray tracing (RT). An adaptive particle refinement (APR) method, compatible with multiphase SPH and optimized for GPU acceleration, is developed to enhance computational efficiency of the multiphase model. The RT model is also optimized and integrated with the APR-GPU architecture, further improving performance. A physics-based wetting force model is introduced, along with a novel method for improving normal vector accuracy near the contact line. The proposed framework is validated through three benchmark cases and applied to simulate LPBF processes. The results demonstrate that the framework achieves high accuracy and efficiency in resolving key LPBF phenomena, including melt pool dynamics and keyhole formation.}, archive = {J_CMAME}, author = {Yibo Ma and Zhilang Zhang and Christian Weißenfels and Minli Zhou and Lingxiao Ma and Xiaofei Tang and Moubin Liu}, doi = {10.1016/j.cma.2025.118423}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118423}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {GPU-accelerated multi-phase, multi-resolution SPH method with ray tracing for laser powder bed fusion}, volume = {448}, year = {2026}, }
• (2026). Physics-guided hybrid network for predicting nonlinear dynamic response of structures under bi-directional ground motions. CMAME, 448, 118422. (www)

  @article{ , abstract = {Seismic structural response is a critical indicator for assessing collapse resistance and ensuring post-earthquake functionality. Therefore, accurate and rapid prediction of these responses is essential, with Deep Learning (DL) models offering a robust alternative to finite element methods due to their computational efficiency and capability to model nonlinear dynamic responses of structures. However, prevailing DL approaches predominantly focus on unidirectional seismic excitations, often neglecting the complex effects of bidirectional seismic excitations on real-world structures. Furthermore, while the development of the physics-informed neural network effectively enhances interpretability under limited data, most existing approaches incorporate physical constraints solely into the loss function, potentially leading to optimization conflicts and slow convergence rates. To bridge these gaps, this study introduces a novel physics-guided hybrid deep learning framework, the Physical Residual Long Short-term Memory network-Kolmogorov–Arnold network model (Phy-RLK), for real-time bidirectional seismic structural response prediction. The proposed DL architecture embeds the Newmark- β numerical integration scheme as a physical residual within LSTM cells. Simultaneously, the adaptive basis functions of the KAN enhance the feature extraction capability from bidirectional seismic inputs, enabling real-time correction of acceleration, velocity, and displacement predictions at each time step, thereby ensuring improved accuracy and physical consistency. The effectiveness of the proposed model is initially verified on a six-story Reinforced Concrete (RC) frame structure, determining optimal hyperparameters in the process. An ablation study further confirms the necessity of both the physics-residual LSTM cell and KAN layer. The model is subsequently implemented on a five-story RC frame structure to verify accuracy and applicability. The experimental results demonstrate that the Phy-RLK provides superior predictive accuracy and robustness, and significantly higher computational efficiency compared to finite element simulations.}, archive = {J_CMAME}, author = {Zheyi Guo and Jun Xu}, doi = {10.1016/j.cma.2025.118422}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118422}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Physics-guided hybrid network for predicting nonlinear dynamic response of structures under bi-directional ground motions}, volume = {448}, year = {2026}, }
• (2026). Sparse data assimilation for under-resolved large-eddy simulations. CMAME, 448, 118421. (www)

  @article{ , abstract = {The need for accurate and fast scale-resolving simulations of fluid flows, where turbulent dispersion is a crucial physical feature, is evident. Large-eddy simulations (LES) are computationally more affordable than direct numerical simulations, but their accuracy depends on sub-grid scale models and the quality of the computational mesh. In order to compensate related errors, a data assimilation approach for LES is devised in this work. The presented method is based on variational assimilation of sparse time-averaged velocity reference data. Working with the time-averaged LES momentum equation allows to employ a stationary discrete adjoint method. Therefore, a stationary corrective force in the unsteady LES momentum equation is iteratively updated within the gradient-based optimization framework in conjunction with the adjoint gradient. After data assimilation, corrected anisotropic Reynolds stresses are inferred from the stationary corrective force. Ultimately, this corrective force that acts on the mean velocity is replaced by a term that scales the velocity fluctuations through nudging of the corrected anisotropic Reynolds stresses. Efficacy of the proposed framework is demonstrated for turbulent flow over periodic hills and around a square cylinder. Coarse meshes are leveraged to further enhance the speed of the optimization procedure. Time- and spanwise-averaged velocity reference data from high-fidelity simulations is taken from the literature. Our results demonstrate that adjoint-based assimilation of averaged velocity enables the optimization of the mean flow, vortex shedding frequency (i. e., Strouhal number), and anisotropic Reynolds stresses. This highlights the superiority of scale-resolving simulations such as LES over simulations based on the (unsteady) Reynolds-averaged equations.}, archive = {J_CMAME}, author = {Justin Plogmann and Oliver Brenner and Patrick Jenny}, doi = {10.1016/j.cma.2025.118421}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118421}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Sparse data assimilation for under-resolved large-eddy simulations}, volume = {448}, year = {2026}, }
• (2026). Improved accuracy of dirichlet-like microscopic solutions on representative volume elements. CMAME, 448, 118420. (www)

  @article{ , abstract = {This work focuses on the improvement of Dirichlet Boundary Value Finite Element solutions on Representative Volume Elements (RVE). For random microstructure Finite Element calculations, it is well-known that Periodic Boundary Conditions (PBC) applied on fixed size periodic RVE are much more precise than classical Dirichlet boundary conditions, so-called Uniform Strain Boundary Conditions (USBC), at the expense of additional efforts. The numerical experiments reported here clearly reveal that the inaccuracy of USBC is mainly due to boundary effects located around and in artificially cut inclusions because of the classical RVE cubic shape. To improve USBC solutions, without affecting their competitive computational speed and memory requirements, existing Dirichlet-based post-processing and pre-processing approaches are benchmarked through an in-depth comparison on various measure of interest: apparent properties, phase first and second moments, frequency histograms. The post-processing methods are namely truncation and filtering. For this latter, a dedicated cut inclusions filter is proposed. The pre-processing methods consist in modifying the inclusions sampling law or the RVE shape to avoid boundary cut inclusions. The main advantage of these pre-processing strategies is to obtain improved local fields, and hence to avoid dedicated post-processing. Innovative Dirichlet-like Boundary Conditions (namely Homogeneous Strain Boundary Conditions - HSBC) based on a rigid body motion of inclusions at the RVE boundary are also introduced. For enough stiff inclusions (material contrast greater than 10), these latter show a promising efficiency on apparent properties but also on local and phase-average fields (even second moments).}, archive = {J_CMAME}, author = {Louis Belgrand and Isabelle Ramière and Marc Josien and Frédéric Lebon}, doi = {10.1016/j.cma.2025.118420}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118420}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Improved accuracy of dirichlet-like microscopic solutions on representative volume elements}, volume = {448}, year = {2026}, }
• (2026). DGTO: Derivable geodesics-coupled topology optimization for multi-axis 3D printing of continuous fiber-reinforced spatial structures. CMAME, 448, 118419. (www)

  @article{ , abstract = {Continuous fiber reinforced composites (CFRCs) are composite materials with exceptional mechanical properties to enhance structural mechanical performance. In comparison with traditional three-axis 3D printing (also referred to as 2.5D printing), multi-axis 3D printing simultaneously moves the nozzle and rotates the build platform during the printing process, making it particularly suited for fabricating spatial structures made of CFRCs due to better alignment between the fiber depositions and principal stress directions. In this research, we propose a Derivable Geodesics-coupled Topology Optimization (DGTO) method for design of CFRCs given the manufacturing scheme of multi-axis 3D printing. A prominent feature of DGTO is the introduction of two geodesic fields to achieve curved layer generation and continuous fiber path planning. The heat diffusion equation and Poisson equation are solved to produce the geodesic fields, and hence, all objective functions and constraints related to the slices and paths are derivable, making them perfectly suitable to be integrated with topology optimization. Hence, the proposed method concurrently optimizes the density field and the auxiliary geodesic fields, simultaneously tuning the material distribution and spatial fiber distribution, thereby attaining optimal performance while fulfilling the manufacturing constraints of multi-axis printing, i.e., self-support of structure and overlap/gap-free of continuous fibers. Four numerical examples are presented to demonstrate the effectiveness of the algorithm, especially showing outstanding performances than designs for traditional three-axis 3D printing.}, archive = {J_CMAME}, author = {Kaixian Liang and Jikai Liu and Shuzhi Xu and Yifan Guo}, doi = {10.1016/j.cma.2025.118419}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118419}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {DGTO: Derivable geodesics-coupled topology optimization for multi-axis 3D printing of continuous fiber-reinforced spatial structures}, volume = {448}, year = {2026}, }
• (2026). Active evolutionary gaussian process for structural large-scale full-field reliability analysis and critical domain prognosis with only few initial samples. CMAME, 448, 118418. (www)

  @article{ , abstract = {Reliability analysis is crucial for ensuring structural integrity, yet it requires repeated, time-consuming evaluations of responses and multivariate limit state functions, while struggling to provide full-field estimations efficiently. Therefore, we propose an active evolutionary reduced Gaussian Process framework ( AER-GP ) for fast and large-scale full-field reliability analysis and critical domain prognosis, utilizing only a few initial samples. In which we first define a novel probability indicator of erroneously evaluating the sign of the minimum of the full-field limit state function. Based on this, we then develop an efficient convergence criterion that relies on the expected error in failure probability estimates. Furthermore, we advance the dual order-reduced Gaussian process coupled Monte Carlo methods to accurately predict the large-scale full-field stochastic response under material and load uncertainty. Where the preliminary design of experiments starts from a significantly small number of sets (e.g.,5), and is progressively added by those samples containing the most information to distinguish the failure boundary. More importantly, we propose a novel mechanism for structural critical domain prognosis mechanism based on the failure probability of each single material point within the entire field, and make accurate critical domain prognosis. Real-world examples, including a car wheel hub and a single-tower cable-stayed bridge, demonstrate that the proposed algorithm achieves high prediction accuracy, computational efficiency, and robustness in large-scale structural reliability analysis and critical domain prognosis. It consistently outperforms widely used methods such as AK-MCS, EFF-MCS, ERF-MCS, and H-MCS. Notably, the proposed AER-GP method requires only a small number of initial DoE samples (fewer than 40), and through adaptive learning, it can reliably produce results with very low errors (typically less than 2%).}, archive = {J_CMAME}, author = {Xinlong Li and Chensen Ding}, doi = {10.1016/j.cma.2025.118418}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118418}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Active evolutionary gaussian process for structural large-scale full-field reliability analysis and critical domain prognosis with only few initial samples}, volume = {448}, year = {2026}, }
• (2026). Mixed-dimensional analysis for coupling 2D elastodynamics and timoshenko beam. CMAME, 448, 118416. (www)

  @article{ , abstract = {Wave propagation is considered in a two-dimensional (2D) elastic structure, which includes a relatively small region whose behavior is fully 2D and a long and slender region whose bending behavior is like that of a Timoshenko beam. To save in computational effort, the latter region is reduced to a genuinely one-dimensional (1D) Timoshenko beam. The mathematical and computational problem posed then involves the coupling of the two regions, such that a well-posed, accurate, numerically stable and efficient hybrid 2D-elastic-Timoshenko-beam model is formed. The appropriate interface conditions are derived, and the well-posedness of the time-dependent problem is proved. A new computational coupling method is proposed, where the shape functions associated with the axial degrees of freedom on the interface of the elastic solid are modified in a special manner, to allow for the rotation continuity. The method results in a symmetric, positive and stable finite element formulation. Numerical examples are presented which demonstrate the performance of the scheme.}, archive = {J_CMAME}, author = {Daniel Rabinovich and Abimael F.D. Loula and Dan Givoli}, doi = {10.1016/j.cma.2025.118416}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118416}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Mixed-dimensional analysis for coupling 2D elastodynamics and timoshenko beam}, volume = {448}, year = {2026}, }
• (2026). Variable stiffness boundary condition to determine effective toughness of heterogeneous materials. CMAME, 448, 118414. (www)

  @article{ , abstract = {We present a computational framework that combines a variable stiffness boundary condition (VSBC) with a phase-field model to evaluate the effective fracture toughness of heterogeneous materials. It is built on the so-called surfing boundary condition (SBC) that applies nonuniform displacement at the remote boundary to stabilize crack-propagation in a heterogeneous medium. Unlike SBC, VSBC is easily implementable in traditional universal testing machines and commercial software packages. The VSBC passively translates a simple, uniform remote displacement into a non-uniform load via an engineered stiffness gradient, enabling the stable, natural propagation of cracks along energetically favorable paths. The framework is validated on homogeneous materials, where the calculated J -integral precisely matches the prescribed fracture toughness. When applied to heterogeneous domains, the VSBC method successfully quantifies the increase in effective toughness due to stiffness and toughness contrasts and captures the critical transition from crack penetration to deflection. Experimental validation using 3D-printed samples confirms the model’s predictive capability. The VSBC framework provides a robust easily accessible tool for investigating fracture in complex materials and guide the design of advanced, fracture-resistant composites.}, archive = {J_CMAME}, author = {Tengyuan Hao and Adrian Piel and Zubaer Hossain}, doi = {10.1016/j.cma.2025.118414}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118414}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Variable stiffness boundary condition to determine effective toughness of heterogeneous materials}, volume = {448}, year = {2026}, }
• (2026). Graph neural network surrogates for contacting deformable bodies with necessary and sufficient contact detection. CMAME, 448, 118413. (www)

  @article{ , abstract = {Surrogate models for the rapid inference of nonlinear boundary value problems in mechanics are helpful in a broad range of engineering applications. However, effective surrogate modeling of applications involving the contact of deformable bodies, especially in the context of varying geometries, is still an open issue. In particular, existing methods are confined to rigid body contact or, at best, contact between rigid and soft objects with well-defined contact planes. Furthermore, they employ contact or collision detection filters that serve as a rapid test but use only the necessary and not sufficient conditions for detection. In this work, we present a graph neural network architecture that utilizes continuous collision detection and, for the first time, incorporates sufficient conditions designed for contact between soft deformable bodies. We test its performance on two benchmarks, including a problem in soft tissue mechanics of predicting the closed state of a bioprosthetic aortic valve. We find a regularizing effect on adding additional contact terms to the loss function, leading to better generalization of the network. These benefits hold for simple contact at similar planes and element normal angles, and complex contact at differing planes and element normal angles. We also demonstrate that the framework can handle varying reference geometries. However, such benefits come with high computational costs during training, resulting in a trade-off that may not always be favorable. We quantify the training cost and the resulting inference speedups on various hardware architectures. Importantly, our graph neural network implementation results in up to a hundred- to thousand-fold speedup on GPU, and twenty- to two hundred-fold speedup on CPU for our benchmark problems at inference.}, archive = {J_CMAME}, author = {Vijay K. Dubey and Collin E. Haese and Osman Gültekin and David Dalton and Manuel K. Rausch and Jan Fuhg}, doi = {10.1016/j.cma.2025.118413}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118413}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Graph neural network surrogates for contacting deformable bodies with necessary and sufficient contact detection}, volume = {448}, year = {2026}, }
• (2026). Personalized multiscale modeling of left atrial mechanics and blood flow. CMAME, 448, 118412. (www)

  @article{ , abstract = {We present a personalized multiscale mechanics model of the left atrium (LA) to simulate its deformation throughout the cardiac cycle and drive blood flow. Our patient data-driven model tightly integrates 3D structural mechanics of the LA myocardium, incorporating both passive and active components, with a 0D closed-loop lumped parameter network (LPN)-based circulatory system model. A finite element (FE) model of LA tissue is constructed from the patient’s images, assuming uniform thickness and employing rule-based fiber directions. We then adopted a multi-step personalization approach, in which the LPN parameters with a surrogate LA model are first optimized to match cuff-based blood pressures and cardiac lumen volumes derived from time-resolved 3D gated computed tomography angiography (CTA) images. The surrogate LA pressure during passive expansion is used to estimate myocardial passive mechanics parameters and the reference unloaded configuration using an inverse finite element analysis (iFEA) framework. Finally, a robust multiscale coupling is applied between the iFEA-optimized FE model and the tuned 0D LPN model to characterize LA contraction. This effectively captures the physiological LA pressure-volume curve and reasonably aligns with the image-based cavity volumes and deformation. We then imposed the resulting simulation-predicted deformation as a moving-wall boundary condition to model atrial hemodynamics. We analyzed the model sensitivities to various simplifications to demonstrate its robustness and versatility and discussed potential future improvements. Overall, this comprehensive digital twinning platform could be applied to study LA biomechanics in health and disease and assist in devising personalized treatment plans.}, archive = {J_CMAME}, author = {Lei Shi and Boyang Gan and Ian Y. Chen and Vijay Vedula}, doi = {10.1016/j.cma.2025.118412}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118412}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Personalized multiscale modeling of left atrial mechanics and blood flow}, volume = {448}, year = {2026}, }
• (2026). Quantum neural network-assisted topology optimization: Concept and implementation with parameterized quantum circuits. CMAME, 448, 118411. (www)

  @article{ , abstract = {This study proposes a quantum machine learning-assisted approach to accelerating density-based topology optimization using quantum neural network (QNN), a class of trainable models based on parameterized quantum circuits (PQCs). The proposed framework extracts key features from classical data, starting by encoding the corresponding finite element analysis results, i.e., strain energy, sensitivity, and design variables from early design iterations obtained using the standard optimizer. Then, the PQCs are fine-tuned using minimization of the binary cross-entropy loss function, enabling the model to learn the mapping between input features and optimal design. Specifically, the framework consists of two main stages. First, an offline training stage where the QNN is calibrated using precomputed iterative results from a relatively coarse mesh obtained through the standard optimizer, establishing pattern recognition between input features and the final design variables. The second is an online stage where the trained QNN model is integrated with the standard optimizer to accelerate the final design. Numerical results show that QNN requires only a small number of qubits, and once trained on coarse meshes with several different boundary conditions, can effectively integrate with standard optimizers to predict target designs across various configurations, including different resolutions, volume constraints, and loading conditions. Furthermore, the proposed QNN-assisted framework significantly reduces computational time compared to standard iterative approaches, laying the groundwork for solving large-scale problems with near-term quantum computers.}, archive = {J_CMAME}, author = {Naruethep Sukulthanasorn and Kenjiro Terada}, doi = {10.1016/j.cma.2025.118411}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118411}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Quantum neural network-assisted topology optimization: Concept and implementation with parameterized quantum circuits}, volume = {448}, year = {2026}, }
• (2026). Generative emulation of chaotic dynamics with coherent prior. CMAME, 448, 118410. (www)

  @article{ , abstract = {Data-driven emulation of nonlinear dynamics is challenging due to long-range skill decay that often produces physically unrealistic outputs. Recent advances in generative modeling aim to address these issues by providing uncertainty quantification and correction. However, the quality of generated simulation remains heavily dependent on the choice of conditioning prior. In this work, we present an efficient generative framework for nonlinear dynamics emulation, connecting principles of turbulence with diffusion-based modeling: Cohesion. Our method estimates large-scale coherent structure of the underlying dynamics as guidance during the denoising process, where small-scale fluctuation in the flow is then resolved. These coherent prior are efficiently approximated using reduced-order models, such as deep Koopman operators, that allow for rapid generation of long prior sequences while maintaining stability over extended forecasting horizon. With this gain, we can reframe forecasting as trajectory planning, a common task in reinforcement learning, where conditional denoising is performed once over entire sequences, minimizing the computational cost of autoregressive-based generative methods. Numerical evaluations on chaotic systems of increasing complexity, including Kolmogorov flow, shallow water equations, and subseasonal-to-seasonal climate dynamics, demonstrate Cohesion superior long-range forecasting skill that can efficiently generate physically-consistent simulations, even in the presence of partially-observed guidance.}, archive = {J_CMAME}, author = {Juan Nathaniel and Pierre Gentine}, doi = {10.1016/j.cma.2025.118410}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118410}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Generative emulation of chaotic dynamics with coherent prior}, volume = {448}, year = {2026}, }
• (2026). Concurrent 3D topology optimization method for hierarchical hybrid structures under static and dynamic loads with CPU-GPU heterogeneous parallelism. CMAME, 448, 118408. (www)

  @article{ , abstract = {Topology optimization of 3D hierarchical hybrid structures (HHS) is constrained by the coupling of high-dimensional design spaces and multiscale computational complexity, often addressed by restricting certain designable components, which limits the full exploration of the design space and realization of performance potential. This paper proposes a novel concurrent topology optimization method for 3D-HHS, achieving concurrent optimization of all designable components, including macroscopic topology, substructural topology, and their spatial distribution, under static and dynamic loads. This approach significantly expands the design space, enhancing the mechanical performance of hierarchical structures. To address the computational challenges of large-scale 3D problems, we employ CPU-GPU heterogeneous parallel computing to improve the efficiency of structural response and sensitivity analysis. Numerical examples demonstrate that this method delivers superior 3D-HHS designs with markedly improved optimization efficiency, providing an innovative solution for efficient 3D structural optimization.}, archive = {J_CMAME}, author = {Yunfei Liu and Ruxin Gao and Ying Li and Daining Fang}, doi = {10.1016/j.cma.2025.118408}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118408}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Concurrent 3D topology optimization method for hierarchical hybrid structures under static and dynamic loads with CPU-GPU heterogeneous parallelism}, volume = {448}, year = {2026}, }
• (2026). Tensegrity structures and data-driven analysis for 3D cell mechanics. CMAME, 448, 118406. (www)

  @article{ , abstract = {The cytoskeleton (CSK) plays an important role in many cell functions. Given the similarities between the mechanical behavior of tensegrity structures and the CSK, many studies have proposed different tensegrity-based models for simulating cell mechanics. However, the low symmetry of most tensegrity units has hindered the analysis of realistic 3D structures. As a result, tensegrity-based modeling in cell mechanics has been mainly focused on single cells or monolayers. In this paper, we propose a 3D tensegrity model based on the finite element method for simulating 3D cell mechanics. We show that the proposed model not only captures the nonlinearity of a single cell in an indentation test and a monolayer in stretch test but also the non-uniform stress distribution in multicellular spheroids upon non-uniform prestress design. Furthermore, we introduce a multiscale data-driven framework for cellular mechanics to optimize the computation, thus paving the way for modeling the mechanobiology of large cellular assemblies such as organs.}, archive = {J_CMAME}, author = {Ziran Zhou and Jacinto Ulloa and Guruswami Ravichandran and José E. Andrade}, doi = {10.1016/j.cma.2025.118406}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118406}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Tensegrity structures and data-driven analysis for 3D cell mechanics}, volume = {448}, year = {2026}, }
• (2026). Configuration-decoupled concurrent topology optimization of heterogeneous lattice structures. CMAME, 448, 118405. (www)

  @article{ , abstract = {Multi-configuration lattice structures have recently been introduced into structural optimization due to their broadly tunable physical properties. Traditional methods for multi-configuration lattice optimization employ extreme strategies of complete fusion or separation, leading to a trade-off between optimality and scalability that has not been fully addressed in the existing literature. The paper suggests decomposing the lattice library into pairs of lattices, through which multi-configuration lattice optimization is decoupled into the concurrent optimization of iso-value, combination category, and ratio within combination. A novel hybrid interpolation scheme is proposed to describe the effective mechanical behavior of the configuration-decoupled lattices. In this approach, polynomial models are employed to characterize the performance of individual lattice combinations, while the Uniform Multiphase Materials Interpolation model is used to integrate the contributions of all combinations. Benchmark experiments, including full-scale simulations, are conducted to validate the effectiveness of the framework. The proposed method enables rapid convergence to configuration layouts that align with the principal stress orientations. Compared to single- and dual-configuration designs, it achieves compliance reductions of 61.0 % and 26.2 %, respectively, approaching the performance of density-based topology optimization. Extended numerical experiments reveal the joint influence of resolution and configuration count on the overall performance. This method achieves a better trade-off between optimality and extensibility, enabling more flexible utilization of large lattice databases in practical engineering fields.}, archive = {J_CMAME}, author = {Xinze Shen and Changdong Zhang and Wenhe Liao and Dawei Li and Tingting Liu}, doi = {10.1016/j.cma.2025.118405}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118405}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Configuration-decoupled concurrent topology optimization of heterogeneous lattice structures}, volume = {448}, year = {2026}, }
• (2026). Intrepid MCMC: Metropolis-hastings with exploration. CMAME, 448, 118402. (www)

  @article{ , abstract = {In engineering examples, one often encounters the need to sample from unnormalized distributions with complex shapes that may also be implicitly defined through a physical or numerical simulation model, making it computationally expensive to evaluate the associated density function. For such cases, MCMC has proven to be an invaluable tool. Random-walk Metropolis Methods (also known as Metropolis-Hastings (MH)), in particular, are highly popular for their simplicity, flexibility, and ease of implementation. However, most MH algorithms suffer from significant limitations when attempting to sample from distributions with multiple modes (particularly disconnected ones). In this paper, we present Intrepid MCMC - a novel MH scheme that utilizes a simple coordinate transformation to significantly improve the mode-finding ability and convergence rate to the target distribution of random-walk Markov chains while retaining most of the simplicity of the vanilla MH paradigm. Through multiple examples, we showcase the improvement in the performance of Intrepid MCMC over vanilla MH for a wide variety of target distribution shapes. We also provide an analysis of the mixing behavior of the Intrepid Markov chain, as well as the efficiency of our algorithm for increasing dimensions. A thorough discussion is presented on the practical implementation of the Intrepid MCMC algorithm. Finally, its utility is highlighted through a Bayesian parameter inference problem for a two-degree-of-freedom oscillator under free vibration.}, archive = {J_CMAME}, author = {Promit Chakroborty and Michael D. Shields}, doi = {10.1016/j.cma.2025.118402}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118402}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Intrepid MCMC: Metropolis-hastings with exploration}, volume = {448}, year = {2026}, }
• (2026). Two-level overlapping additive schwarz preconditioner for training scientific machine learning applications. CMAME, 448, 118400. (www)

  @article{ , abstract = {We introduce a novel two-level overlapping additive Schwarz preconditioner for accelerating the training of scientific machine learning applications. The design of the proposed preconditioner is motivated by the nonlinear two-level overlapping additive Schwarz preconditioner. The neural network parameters are decomposed into groups (subdomains) with overlapping regions. In addition, the network’s feed-forward structure is indirectly imposed through a novel subdomain-wise synchronization strategy and a coarse-level training step. Through a series of numerical experiments, which consider physics-informed neural networks and operator learning approaches, we demonstrate that the proposed two-level preconditioner significantly speeds up the convergence of the standard (LBFGS) optimizer while also yielding more accurate machine learning models. Moreover, the devised preconditioner is designed to take advantage of model-parallel computations, which can further reduce the training time.}, archive = {J_CMAME}, author = {Youngkyu Lee and Alena Kopaničáková and George Em Karniadakis}, doi = {10.1016/j.cma.2025.118400}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118400}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Two-level overlapping additive schwarz preconditioner for training scientific machine learning applications}, volume = {448}, year = {2026}, }
• (2026). A cohesive zone treatment for the material point method involving problems of large deformation and damage. CMAME, 448, 118399. (www)

  @article{ , abstract = {A new algorithm is described that permits the use of cohesive zones in the material point method for problems involving large deformation and fracture. In contrast to previous cohesive zone implementations, this method does not utilize massless surface-element particles. Instead, cohesive tractions are computed using the shape function mappings from a reference grid configuration in combination with explicitly defined particle surface normals and surface positions. These normals and relative surface positions are updated each time step according to particle deformation. The tractions are converted to cohesive forces using the nodal areas and mapped back to particles using the same reference shape function mappings. These forces are then remapped by conventional particle-to-grid interpolation as external forces using the current-configuration shape-function mappings. This allows highly compliant cohesive zones to function over jump displacements larger than a grid cell. Upon damage, these interfaces can revert to conventional multi-field contact surfaces. This approach is general and readily applies to two and three dimensions as well as being compatible with damage-field gradient partitioning offering exceptional computational flexibility. The framework for this method enables other capabilities, such as improved contact precision using explicitly defined surface normals and positions, and a method to mitigate spurious material damage at weak discontinuities between stiff brittle materials and soft or compliant materials.}, archive = {J_CMAME}, author = {Cameron M. Crook and Michael A. Homel}, doi = {10.1016/j.cma.2025.118399}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118399}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {A cohesive zone treatment for the material point method involving problems of large deformation and damage}, volume = {448}, year = {2026}, }
• (2026). Overlapping schwarz preconditioners for isogeometric discretizations of acoustic wave problems. CMAME, 448, 118397. (www)

  @article{ , abstract = {The aim of this work is to construct and analyze two-level overlapping additive Schwarz (OAS) preconditioners for isogeometric discretizations of the acoustic wave equation with absorbing boundary conditions. Both Collocation and Galerkin isogeometric methods are employed for space discretization, while time advancing is performed by means of a Newmark implicit scheme. The linear systems to be solved at each time step are ill conditioned, especially in case of highly regular splines, thus their solution requires the use of effective preconditioners. Two-level OAS solvers consist of partitioning the domain into overlapping subdomains, solving independent local problems on each subdomain and an additional coarse problem associated with the subdomain mesh. Several two-dimensional numerical results validate our theoretical estimates, showing the scalability and quasi-optimality of the algorithms proposed. We also investigate numerically the robustness of the OAS preconditioners with respect to the spline polynomial degree, the spline regularity and the overlap parameter.}, archive = {J_CMAME}, author = {Elena Zampieri and Simone Scacchi and Luca F. Pavarino}, doi = {10.1016/j.cma.2025.118397}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118397}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Overlapping schwarz preconditioners for isogeometric discretizations of acoustic wave problems}, volume = {448}, year = {2026}, }
• (2026). Latent space modeling of parametric and time-dependent PDEs using neural ODEs. CMAME, 448, 118394. (www)

  @article{ , abstract = {Partial Differential Equations (PDEs) are central to science and engineering. Since solving them is computationally expensive, a lot of effort has been put into approximating their solution operator via both traditional and recently increasingly Deep Learning (DL) techniques. In this paper, we propose an autoregressive and data-driven method using the analogy with classical numerical solvers for time-dependent, parametric and (typically) nonlinear PDEs. We present how Dimensionality Reduction (DR) can be coupled with Neural Ordinary Differential Equations (NODEs) in order to learn the solution operator of arbitrary PDEs accounting both for (continuous) time and parameter dependency. The idea of our work is that it is possible to map the high-fidelity (i.e., high-dimensional) PDE solution space into a reduced (low-dimensional) space, which subsequently exhibits dynamics governed by a (latent) Ordinary Differential Equation (ODE). Solving this (easier) ODE in the reduced space allows avoiding solving the PDE in the high-dimensional solution space, thus decreasing the computational burden for repeated calculations for e.g., uncertainty quantification or design optimization purposes. The main outcome of this work is the importance of exploiting DR as opposed to the recent trend of building large and complex architectures: we show that by leveraging DR we can deliver not only more accurate predictions, but also a considerably lighter and faster DL model compared to existing methodologies.}, archive = {J_CMAME}, author = {Alessandro Longhi and Danny Lathouwers and Zoltán Perkó}, doi = {10.1016/j.cma.2025.118394}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118394}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Latent space modeling of parametric and time-dependent PDEs using neural ODEs}, volume = {448}, year = {2026}, }
• (2026). Geometry-agnostic model reduction with GNN-generated reduced POD bases and boosted PGD enrichment for (non)linear structural elastodynamics. CMAME, 448, 118357. (www)

  @article{ , abstract = {This contribution proposes a new and significantly enhanced extension of a recently-introduced hybrid Graph Neural Network (GNN)-based reduced-order modeling approach for the numerical solution of time-dependent partial differential equations on non-parametric finite element meshes. Building upon previous proof-of-concept work, this more generalized framework presents a number of key novelties: tight integration of graph-based learning with physical information via direct imposition of finite element operators as node and edge level features; introduction of a Grassmannian subspace distance measure as a dedicated training objective; incorporation of a Gated Recurrent Unit (GRU) for a more efficient and lightweight architecture; hybridization with other Galerkin-based reduced-order methods such as the Proper Orthogonal Decomposition (POD); and a first treatment of nonlinear problems. A novel, on-the-fly enrichment mechanism, modified from a classical Proper General Decomposition (PGD) and dubbed ”Boosted PGD”, is additionally introduced to improve prediction accuracy at low computational cost via additional greedy corrective modes. The efficacy of the overall methodology is assessed on two challenging datasets featuring significant geometric and topological variations that include highly heterogeneous spatial discretizations. A variety of performance studies demonstrate very competitive accuracy and computational cost in simulating highly-dynamic behavior when compared to conventional full-order finite element models, including a remarkable capacity to generalize to configurations well outside of the topological scope of the original training and validation sets. Results imply that solvers constructed from such an approach may enable more scalable and robust mechanical simulations for complex, real-world engineering applications related to iterative design.}, archive = {J_CMAME}, author = {Victor Matray and David Néron and Frédéric Feyel and Faisal Amlani}, doi = {10.1016/j.cma.2025.118357}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118357}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Geometry-agnostic model reduction with GNN-generated reduced POD bases and boosted PGD enrichment for (non)linear structural elastodynamics}, volume = {448}, year = {2026}, }
• (2026). Weighted eigenseparation-based residual approach for model reduction of interface failure in heterogeneous materials. CMAME, 448, 118352. (www)

  @article{ , abstract = {Interface failure plays a critical role in the degradation of heterogeneous materials, often governing structural integrity across a range of applications from fiber-reinforced composites to polycrystalline rocks. This paper introduces a novel model reduction framework—the Weighted Eigenseparation-based Residual (WER) approach—for efficiently simulating interface failure using cohesive zone models. Two variants of the WER are developed: a force-based formulation that weakly enforces equilibrium equations at the interface at the modal level and a separation-based formulation that weakly enforces contact conditions at the modal level. These formulations are supported by precomputed influence functions within representative volume elements (RVEs), significantly reducing computational cost while preserving accuracy. The separation-based variant, in particular, demonstrates broad applicability across microstructures with interface junctions. Numerical examples in two and three dimensions—including fiber composites and geological microstructures—demonstrate the effectiveness, convergence behavior, and computational advantages of the WER over direct numerical simulations. The results show that high-fidelity predictions can be obtained even with coarsely discretized interface modes, confirming the robustness and versatility of the WER methodology.}, archive = {J_CMAME}, author = {Jacob Fish and Junhe Cui}, doi = {10.1016/j.cma.2025.118352}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118352}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Weighted eigenseparation-based residual approach for model reduction of interface failure in heterogeneous materials}, volume = {448}, year = {2026}, }
• (2026). Stretch-based hyperelastic electromechanical constitutive metamodels via gradient enhanced gaussian predictors using hierarchical structure discovery. CMAME, 448, 118349. (www)

  @article{ , abstract = {This paper introduces a new approach to developing electromechanical constitutive metamodels via the use of Gradient Enhanced Gaussian Predictors (Kriging). The formulation uses principal stretches for the isotropic mechanics, invariants for the electrostatics and coupling terms, and accounts for anisotropy through the relevant inclusion of anisotropic invariants associated with a respective symmetry integrity basis. Three novelties are presented in this paper. The first is the use of orthogonal projections to identify the most appropriate set of inputs - related to material anisotropy - for use in the metamodel. By projecting the stress and electric field data into several derivative bases - defined for each anisotropic class - and then reconstructing the quantities, the errors in reconstruction can be assessed thus inferring the most appropriate class of anisotropy. Furthermore, the procedure forms a pre-processing stage and is particularly useful when an underlying model is completely unknown as seen when modelling Relative Volume Elements. The second novelty arises from the use of a hybrid formulation, namely the principal stretches for isotropic mechanics and the electromechanical anisotropic invariants. This is beneficial during the projection procedure in reducing the cases where the projection matrix becomes singular but requires careful development of the correlation function to maintain physical symmetry conditions. Thirdly, the electromechanical metamodels are calibrated upon the concentric styled data before being integrated within a Finite Element framework and tested upon a range of challenging simulations including bending actuators with induced torsion, frilling due to bending with selected electrode placement, as well as buckling plates tested with three rank-one laminate materials with increasing levels of anisotropy due to physical contrasts. The successful calibration and implementation of the metamodels can be witnessed amongst the wide range of presented numerical examples.}, archive = {J_CMAME}, author = {Nathan Ellmer and Rogelio Ortigosa and Jesús Martínez-Frutos and Roman Poya and Johann Sienz and Antonio J. Gil}, doi = {10.1016/j.cma.2025.118349}, journal = {Computer Methods in Applied Mechanics and Engineering}, month = {1}, pages = {118349}, shortjournal = {Comput. Meth. Appl. Mech. Eng.}, title = {Stretch-based hyperelastic electromechanical constitutive metamodels via gradient enhanced gaussian predictors using hierarchical structure discovery}, volume = {448}, year = {2026}, }

COMCOM - 11

• (2025). An effective method for data aggregation point placement in LoRaWAN network for smart metering service. COMCOM, 243, 108335. (www)

  @article{ , abstract = {Electric grids have been restructured with Smart Grids (SGs), and the deployment of Advanced Metering Infrastructure (AMI) systems is a fundamental part of this process. An AMI system consists of Smart Meters (SMs) that collect energy consumption data and send it to the utility company through Data Aggregation Points (DAPs). Thus, methods to determine the appropriate quantity and positions of DAPs become necessary. In this context, this paper proposes a method called DAP Placement (DPlace), which determines the minimum number of DAPs and applies the Fuzzy C-Means algorithm to position the DAPs at coordinates that enhance the successful reception of packets, ensuring that smart metering applications can adequately transmit data through a Long Range Wide-Area Network (LoRaWAN) network. The proposed method is compared to three state-of-the-art solutions through simulations, and the results show that the DPlace method reduces the required number of DAPs by up to 37.04% while achieving communication performance – evaluated through metrics such as packet delivery ratio and delay – similar to related methods, even with less DAPs.}, archive = {J_COMCOM}, author = {Thiago A.R. da Silva and Geraldo A. Sarmento N. and Luís H. de O. Mendes and Pedro F. de Abreu and Artur F. da S. Veloso and Fernando J.V. Santos and José Valdemir dos R. Júnior}, doi = {10.1016/j.comcom.2025.108335}, journal = {Computer Communications}, month = {11}, pages = {108335}, shortjournal = {Comput. Commun.}, title = {An effective method for data aggregation point placement in LoRaWAN network for smart metering service}, volume = {243}, year = {2025}, }
• (2025). META: Multi-classified encrypted traffic anomaly detection with fine-grained flow and interaction analysis. COMCOM, 243, 108333. (www)

  @article{ , abstract = {The pervasive implementation of encryption mechanisms has introduced considerable obstacles to anomalous traffic detection, rendering conventional attack detection methodologies that rely on packet payload characteristics ineffectual. In the absence of plaintext information, current anomaly encrypted traffic detection mainly relies on traffic data analysis to identify and characterize anomalous attack patterns in encrypted traffic, employing machine learning or deep learning models. However, the existing methods still suffer from limited detection capabilities, especially the ability to classify multi-class attacks due to insufficient internal and external features. In this paper, we propose a Multi-classified Encrypted Traffic Anomaly Detection (META) method. META refines and extends the available feature dimensions in encrypted traffic by leveraging two key aspects: the internal interaction behavior information within the traffic and the external interaction behavior information in network topology. Specifically, an in-depth examination of the internal packet interaction features is undertaken, resulting in a novel feature set, designated as META-Features, encompassing 278 fine-grained statistical features. Furthermore, a Graph Neural Network (GNN) is employed to learn the external interaction behavior in the network topology from the embedding of the IP node graph and flow edge graph. The results of the experiments demonstrate that the refined feature set META-Features significantly enhances the model’s detection capabilities. Thereby, the META-GNN model exhibits superior performance compared to the traditional approaches, with an accuracy of 91.90% and an F1-score of 87.41%.}, archive = {J_COMCOM}, author = {Boyu Kuang and Yuchi Chen and Yansong Gao and Yaqian Xu and Anmin Fu and Willy Susilo}, doi = {10.1016/j.comcom.2025.108333}, journal = {Computer Communications}, month = {11}, pages = {108333}, shortjournal = {Comput. Commun.}, title = {META: Multi-classified encrypted traffic anomaly detection with fine-grained flow and interaction analysis}, volume = {243}, year = {2025}, }
• (2025). Evaluating scalability of median-based ADR under different mobility conditions. COMCOM, 243, 108322. (www)

  @article{ , abstract = {The LoRaWAN protocol is widely used in Internet of Things (IoT) applications due to its ability to provide long-range, low-power communication. The Adaptive Data Rate (ADR) mechanism dynamically adjusts transmission parameters to optimize energy consumption. However, ADR is primarily designed for static devices, which limits its effectiveness in mobile environments, where fluctuating signal conditions can degrade performance. To address this limitation, the Median-Based ADR (MB-ADR) scheme was introduced, leveraging statistical measures to improve ADR adaptability to changing channel conditions. This study evaluates the scalability of MB-ADR in networks with up to 1,000 end devices and node speeds of up to 20 m/s, considering mobility models such as Random Walk and Gauss–Markov. The results show that MB-ADR demonstrates superior performance in scenarios with realistic mobility patterns, particularly under the Steady-State Random Waypoint model, resulting in improvements of up to 15% in Packet Delivery Ratio (PDR) and 55% in energy efficiency compared to a Kalman filter-based scheme under the same mobility model. Additionally, the analysis demonstrates the effectiveness of MB-ADR in improving throughput and reducing collisions by promoting an efficient distribution of spreading factors. Overall, the study confirms the potential of MB-ADR to enhance communication reliability and energy efficiency in mobile IoT networks, making it a viable solution for large-scale, high-density IoT deployments with variable mobility.}, archive = {J_COMCOM}, author = {Geraldo A. Sarmento Neto and Thiago A.R. da Silva and Artur F. da S. Veloso and Pedro Felipe de Abreu and Luis H. de O. Mendes and J. Valdemir dos Reis Jr}, doi = {10.1016/j.comcom.2025.108322}, journal = {Computer Communications}, month = {11}, pages = {108322}, shortjournal = {Comput. Commun.}, title = {Evaluating scalability of median-based ADR under different mobility conditions}, volume = {243}, year = {2025}, }
• (2025). Revisiting the problem of optimizing spreading factor allocations in LoRaWAN: From theory to practice. COMCOM, 243, 108321. (www)

  @article{ , abstract = {This paper revisits the problem of optimizing LoRa network success probability by proposing an optimized allocation strategy for Spreading Factors (SFs) under both uniform and Gaussian network deployments with a single or multiple gateways. More specifically, we solve the problem of finding the best SF allocations in dense network deployments whose EDs are first assigned with the minimum SF. Theoretical models are developed to quantify the success probability of transmissions, considering the capture effect as well as intra- and inter-SF interference. A mathematical optimization framework is introduced to determine the optimal SF distribution that maximizes the average probability of packet reception. The problem is solved using Mixed Integer Linear Programming (MILP), and then evaluated using simulations. Even though optimal SF allocation strategies have been proposed in the literature, no practical insights have been discovered and no real-world deployments have been considered. To this extent, the practical benefits of using improved or optimal SF settings are discovered in this paper. Simulation results confirm the theoretical findings while they demonstrate an up to 10 percentage points improvements in Packet Reception Ratio (PRR) in the real-world use-case.}, archive = {J_COMCOM}, author = {Dimitrios Zorbas and Aruzhan Sabyrbek and Luigi Di Puglia Pugliese}, doi = {10.1016/j.comcom.2025.108321}, journal = {Computer Communications}, month = {11}, pages = {108321}, shortjournal = {Comput. Commun.}, title = {Revisiting the problem of optimizing spreading factor allocations in LoRaWAN: From theory to practice}, volume = {243}, year = {2025}, }
• (2025). Energy consumption optimization in UAV-assisted multi-layer mobile edge computing with active transmissive RIS. COMCOM, 243, 108320. (www)

  @article{ , abstract = {Unmanned Aerial Vehicle (UAV)-assisted edge computing provides low-latency and low-energy consumption computing capabilities for sparsely distributed Internet of Things (IoT) networks. In addition, the assisted UAVs provide line-of-sight links to further improve communication quality. However, the existing offloading strategies have low efficiency and high costs. Motivated by this, we propose a novel UAV-assisted multi-layer mobile edge computing network with active transmissive reconfigurable intelligent surface (RIS). The introduced an active transmissive RIS not only receives data from UAVs but also performs computing functionality. We establish an optimization to minimize the total system energy consumption under delay constraints by jointly planning UAV positions and allocating computing bits, sub-carriers, time slots, transmission power, and RIS transmission coefficient. To tackle this problem, we first use the block coordinate descent (BCD) algorithm to decouple it into four sub-problems. Then, we solve them by adopting successive convex approximation (SCA), difference-convex (DC) programming, and introducing slack variables. Experimental results demonstrate that the proposed network is superior to the other five baselines concerning energy consumption reduction. Also, the influences of system parameters are verified, including the number of IoT devices, the number of RIS elements, and the delay threshold.}, archive = {J_COMCOM}, author = {Kexin Yang and Yaxi Liu and Boxin He and Jiahao Huo and Wei Huangfu}, doi = {10.1016/j.comcom.2025.108320}, journal = {Computer Communications}, month = {11}, pages = {108320}, shortjournal = {Comput. Commun.}, title = {Energy consumption optimization in UAV-assisted multi-layer mobile edge computing with active transmissive RIS}, volume = {243}, year = {2025}, }
• (2025). RIS-assisted LoRa networks with diversity: Impact of hardware impairments and phase noise. COMCOM, 243, 108319. (www)

  @article{ , abstract = {This paper investigates the performance of downlink LoRa networks assisted by reconfigurable intelligent surfaces (RIS) and diversity techniques. We derive closed-form expressions for the coverage probability (Pcov) under four scenarios: phase noise at the RIS only, hardware impairments at both the gateway and end devices (EDs), the combined effect of both impairments, and an ideal benchmark case. The analysis is carried out within a unified framework that is valid for any number of RIS elements, providing key insights into the influence of hardware impairment levels, gateway transmit power, and the diversity order as the number of RIS elements grows large. The results reveal that coverage probability improves with transmit power but deteriorates under more severe hardware impairments, while the diversity order scales directly with the number of RIS elements. Monte Carlo simulations validate the analytical findings and confirm that the ideal scenario achieves the best performance, followed in order by the phase noise, hardware impairment, and combined impairment cases.}, archive = {J_COMCOM}, author = {Thi-Phuong-Anh Hoang and Thien Huynh-The and Tien Hoa Nguyen and Trong-Thua Huynh and Nguyen-Son Vo and Lam-Thanh Tu}, doi = {10.1016/j.comcom.2025.108319}, journal = {Computer Communications}, month = {11}, pages = {108319}, shortjournal = {Comput. Commun.}, title = {RIS-assisted LoRa networks with diversity: Impact of hardware impairments and phase noise}, volume = {243}, year = {2025}, }
• (2025). Channel-hopping sequence generation for blind rendezvous in cognitive radio-enabled internet of vehicles: A multi-agent twin delayed deep deterministic policy gradient-based method. COMCOM, 243, 108318. (www)

  @article{ , abstract = {Efficient spectrum utilization is a major challenge in highly dynamic vehicular environments due to the scarcity of spectrum resources. Cognitive Radio (CR) has emerged as a solution to improve spectrum utilization by enabling opportunistic access in IoV. In this context, channel-hopping based blind rendezvous offers a practical approach for decentralized spectrum access in CR-enabled IoV (CR-IoV). This paper presents a novel Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3PG)-based strategy for generating channel sequences in channel-hopping-based blind rendezvous. Unlike existing methods that overlook the quality of licensed spectrum, our approach ensures spectrum efficiency and QoS awareness in dynamic channel sequence generation. We formulate the channel sequence selection problem as a multi-objective optimization, aiming to maximize spectrum efficiency and minimize Time-To-Rendezvous (TTR) while meeting stringent latency and reliability requirements for vehicular communications. Each vehicle independently generates a channel-hopping sequence using a learning agent, which considers key channel quality metrics such as availability, reliability, and capacity. The generated sequences are employed in an asynchronous and asymmetric blind rendezvous process, enhancing adaptability to dynamic network conditions. Simulation results demonstrate that the proposed method significantly outperforms existing approaches, including Enhanced Jump-Stay (EJS), Single-radio Sunflower Set (SSS), Zero-type, One-type, and S-type (ZOS), Multi-Agent Q-Learning based Rendezvous (MAQLR), Exponential-weight algorithm for Exploration and Exploitation (Exp3), and Reinforcement Learning-based Channel-Hopping Rendezvous (RLCH) in terms of Expected TTR (ETTR), Maximum TTR (MTTR), delay, capacity, and reliability.}, archive = {J_COMCOM}, author = {Mehri Asadi Vasfi and Behrouz Shahgholi Ghahfarokhi}, doi = {10.1016/j.comcom.2025.108318}, journal = {Computer Communications}, month = {11}, pages = {108318}, shortjournal = {Comput. Commun.}, title = {Channel-hopping sequence generation for blind rendezvous in cognitive radio-enabled internet of vehicles: A multi-agent twin delayed deep deterministic policy gradient-based method}, volume = {243}, year = {2025}, }
• (2025). Tensor-based task allocation using multi-objective optimization in GECC environment. COMCOM, 243, 108316. (www)

  @article{ , abstract = {Green Edge-Cloud Computing (GECC) has emerged as a promising paradigm to meet the diverse requirements of modern applications by integrating edge and cloud resources. Existing task allocation strategies in GECC environment often fail to adequately address the problems of low resource utilization and high economic cost in multi-objective conflicts. Therefore, this paper proposes a tensor-based task allocation scheme using multi-objective optimization in GECC environment. We first extend the task allocation problem in GECC environment to a multi-objective optimization problem and construct five optimization models, i.e., energy, system reliability, quality of experience, economic cost, and latency. Then, to address the complex relationship among these objectives, we develop a tensor-based representation and calculation model for task allocation across cloud, edge service, and edge device platforms. Furthermore, we propose a tensor-based multi-objective beetle swarm optimization algorithm combined speed limiting and dynamic step strategies (TMOBSO-SLDS) that dynamically adjusts the step size and limit speed to improve the global search efficiency and the diversity of solution set. Extensive experimental results in various task allocation scenarios demonstrate that our proposed TMOBSO-SLDS algorithm outperforms existing approaches, as measured by the HV value. It can significantly enhance the diversity of the solution set and improve resource utilization.}, archive = {J_COMCOM}, author = {Huazhong Liu and Longtao Huang and Yuan Tian and Jihong Ding and Xiaoxue Yin and Guangshun Zhang}, doi = {10.1016/j.comcom.2025.108316}, journal = {Computer Communications}, month = {11}, pages = {108316}, shortjournal = {Comput. Commun.}, title = {Tensor-based task allocation using multi-objective optimization in GECC environment}, volume = {243}, year = {2025}, }
• (2025). A preemptive task unloading scheme based on second optional unloading in cloud-fog collaborative networks. COMCOM, 243, 108315. (www)

  @article{ , abstract = {Long-distance data transmission between Internet of Things (IoT) devices and remote cloud center often leads to unacceptable latency for certain tasks. Fog computing has emerged as a promising solution for low-latency tasks. Consequently, the concept of cloud-fog collaborative networks has garnered significant attention. However, existing research primarily focuses on heterogeneous tasks, overlooking the crucial aspect of considering both task priority and second unloading. To address this gap, this paper proposes a novel task unloading scheme that concurrently takes preemptive priority and second optional unloading into account. In this scheme, delay-sensitive tasks (DSTs) are given preemptive priority over delay-tolerant tasks (DTTs). Furthermore, some DTTs may undergo preprocessing in the fog layer to optimize resource utilization. Moreover, tasks encountering blocking or preemption in the fog layer can also be secondarily unloaded to the cloud layer. In this framework, we devise a four-dimensional Markov chain (4DMC) to model and analyze this process. Through numerical experiments, we assess performance indicators under various parameters. Ultimately, our proposed strategy is compared with the unloading scheme that does not incorporate second unloading through both numerical analysis and simulation validation. The results indicate that our scheme notably enhances the throughput of DTTs, albeit at a marginal performance trade-off.}, archive = {J_COMCOM}, author = {Yuan Zhao and Hongmin Gao and Shuaihua Liu}, doi = {10.1016/j.comcom.2025.108315}, journal = {Computer Communications}, month = {11}, pages = {108315}, shortjournal = {Comput. Commun.}, title = {A preemptive task unloading scheme based on second optional unloading in cloud-fog collaborative networks}, volume = {243}, year = {2025}, }
• (2025). ScaleIP: A hybrid autoscaling of VoIP services based on deep reinforcement learning. COMCOM, 243, 108314. (www)

  @article{ , abstract = {Adaptive resource provisioning has become crucial for cloud-based applications, especially those managing real-time traffic like Voice over IP (VoIP), which experience rapidly fluctuating workloads. Traditional static provisioning methods often fall short in these dynamic environments, leading to inefficiencies and potential service disruptions. Existing solutions struggle to maintain performance under varying traffic conditions, particularly for time-sensitive applications. This paper introduces ScaleIP, a hybrid autoscaling solution for containerized VoIP services that offers real-time adaptability and efficient resource management. ScaleIP leverages Deep Reinforcement Learning to make dynamic and efficient scaling decisions, improving call latency, increasing the number of successfully routed calls, and maximizing resource utilization. We evaluated ScaleIP through extensive experiments conducted on a real testbed utilizing the customer Call Detail Record (CDR) from 2023 provided by World Direct, encompassing over 89 million calls. The results show that ScaleIP consistently maintains call latency below 2 s, increases the number of successfully routed calls by 3.26 ×, and increases the resource utilization up to 60 % compared to state-of-the-art autoscaling methods.}, archive = {J_COMCOM}, author = {Zahra Najafabadi Samani and Juan Aznar Poveda and Dominik Gratz and Rene Hueber and Philipp Kalb and Thomas Fahringer}, doi = {10.1016/j.comcom.2025.108314}, journal = {Computer Communications}, month = {11}, pages = {108314}, shortjournal = {Comput. Commun.}, title = {ScaleIP: A hybrid autoscaling of VoIP services based on deep reinforcement learning}, volume = {243}, year = {2025}, }
• (2025). Beyond performance comparing the costs of applying deep and shallow learning. COMCOM, 243, 108312. (www)

  @article{ , abstract = {The rapid growth of mobile network traffic and the emergence of complex applications, such as self-driving cars and augmented reality, demand ultra-low latency, high throughput, and massive device connectivity, which traditional network design approaches struggle to meet. These issues were initially addressed in Fifth-Generation (5G) and Beyond-5G (B5G) networks, where Artificial Intelligence (AI), particularly Deep Learning (DL), is proposed to optimize the network and to meet these demanding requirements. However, the resource constraints and time limitations inherent in telecommunication networks raise questions about the practicality of deploying large Deep Neural Networks (DNNs) in these contexts. This paper analyzes the costs of implementing DNNs by comparing them with shallow ML models across multiple datasets and evaluating factors such as execution time and model interpretability. Our findings demonstrate that shallow ML models offer comparable performance to DNNs, with significantly reduced training and inference times, achieving up to 90% acceleration. Moreover, shallow models are more interpretable, as explainability metrics struggle to agree on feature importance values even for high-performing DNNs.}, archive = {J_COMCOM}, author = {Rafael Teixeira and Leonardo Almeida and Pedro Rodrigues and Mário Antunes and Diogo Gomes and Rui L. Aguiar}, doi = {10.1016/j.comcom.2025.108312}, journal = {Computer Communications}, month = {11}, pages = {108312}, shortjournal = {Comput. Commun.}, title = {Beyond performance comparing the costs of applying deep and shallow learning}, volume = {243}, year = {2025}, }

COR - 41

• (2026). The obnoxious facilities p-center problem with forbidden regions. COR, 185, 107297. (www)

  @article{ , abstract = {We formulate a planar obnoxious p -center problem with circular forbidden regions, such as a minimum required distance from some or all the demand points. We use geometric computation to develop a tight lower bound on the optimal value of the objective function and propose solution techniques which yield optimal solutions to several multiple large-scale tsplib instances in less than few hours’ time. We observe that, for larger forbidden regions, the obnoxious p -center problem reaches the best possible overall objective for a relatively small number of facilities. Our proposed solution approach is particularly useful in instances with multiple overlapping forbidden regions.}, archive = {J_COR}, author = {Pawel Kalczynski and Zvi Drezner}, doi = {10.1016/j.cor.2025.107297}, journal = {Computers & Operations Research}, month = {1}, pages = {107297}, shortjournal = {Comput. Oper. Res.}, title = {The obnoxious facilities p-center problem with forbidden regions}, volume = {185}, year = {2026}, }
• (2026). An adaptive and efficient path planning algorithm for UAV navigation in complex environments. COR, 185, 107296. (www)

  @article{ , abstract = {Efficient and reliable path planning is critical for autonomous unmanned aerial vehicle (UAV) navigation in complex environments, including urban obstacle fields and sensor-restricted airspaces. This paper proposes the rapidly converging and greedy-optimized RRT* (RGO-RRT*) algorithm, which integrates four adaptive modules to enhance planning performance. These include: (1) region-based probabilistic sampling that prioritizes high-potential regions to reduce redundant exploration and accelerate convergence; (2) adaptive step-size adjustment based on obstacle density for fine maneuverability in cluttered areas and rapid expansion in open spaces; (3) dynamic goal biasing that gradually increases goal attraction to balance exploration and convergence; and (4) an improved artificial potential field with an adaptive repulsion model to mitigate local minima and ensure smoother trajectories. Additionally, three auxiliary strategies — bidirectional tree expansion, greedy optimization, and feasibility constraints-are employed to further refine path quality and search efficiency. Extensive simulations are conducted in four structurally diverse environments to evaluate performance under various levels of obstacle density and geometric complexity. Results show that RGO-RRT* consistently outperforms five benchmark algorithms (RRT*, Bi-RRT*, APF-RRT*, Bi-APF-RRT*, and Improved Bi-APF-RRT*), achieving up to 83.1% fewer iterations, 11.7% shorter path lengths, and 87.9% reduction in planning time. These findings demonstrate the method’s robustness, efficiency, and applicability to UAV navigation in cluttered scenarios.}, archive = {J_COR}, author = {Yonggang Wang and Guoliang Wang}, doi = {10.1016/j.cor.2025.107296}, journal = {Computers & Operations Research}, month = {1}, pages = {107296}, shortjournal = {Comput. Oper. Res.}, title = {An adaptive and efficient path planning algorithm for UAV navigation in complex environments}, volume = {185}, year = {2026}, }
• (2026). On-demand meal delivery: Drone scheduling with battery replacement optimization. COR, 185, 107295. (www)

  @article{ , abstract = {Drones have become a promising solution for on-demand delivery thanks to their ability to travel fast and navigate without road restrictions. In the context of direct meal delivery using drones, it is a common practice to replace the drone’s battery after each round trip between the central launch site and customer locations to prevent power interruptions. This practice can lead to frequent battery replacements, resulting in increased downtime and decreased drone utilization. To enhance the efficiency of drone delivery, we take into account load-dependent energy consumption for the drones and optimize battery replacement along with drone scheduling. A mixed integer programming formulation is constructed to mathematically capture the problem. Additionally, we develop a time-expanded network flow method and a tailored hybrid variable neighborhood search algorithm to solve the problem exactly and heuristically. Computational studies validate the effectiveness and efficiency of the proposed operational model and solution approaches. The results indicate that optimizing battery replacement can induce an increase in on-time deliveries by up to 7.69% compared to replacing batteries after each return, and by 14.29% compared to only replacing them when energy levels are low. Such benefits are particularly significant in scenarios with tighter delivery deadlines and longer battery replacement times.}, archive = {J_COR}, author = {Wenqian Liu and Yandong He and Ginger Y. Ke and Lianmin Zhang}, doi = {10.1016/j.cor.2025.107295}, journal = {Computers & Operations Research}, month = {1}, pages = {107295}, shortjournal = {Comput. Oper. Res.}, title = {On-demand meal delivery: Drone scheduling with battery replacement optimization}, volume = {185}, year = {2026}, }
• (2026). Optimizing production–transportation–delivery in global supply chain with demand ambiguity by branch-and-cut algorithm. COR, 185, 107293. (www)

  @article{ , abstract = {The complexity incorporated in global supply chain (GSC) means the production, transportation, and delivery are totally operating and completing in the dynamic business environment with unforeseen events. At present, there are two key challenges in the transnational supply chain network: addressing the demand ambiguity and enhancing cooperation among supply chain entities. To optimize the production–transportation–delivery decision in GSC, a novel globalized distributionally robust GSC (GDR-GSC) model with horizontal cooperation is proposed, in which the ambiguity of demand distribution is characterized by inner and outer ambiguity sets. Subsequently, the proposed model is transformed into mixed integer nonlinear programming (MINLP) model by duality theory. It is commonly difficult to solve in high-dimensional case. Therefore, a customized Branch-and-Cut (B&C) algorithm tailored for the GDR-GSC model is designed to handle complex MINLP problems, and improves computational efficiency and solution quality. The case study based on Apple’s sales operations in China and Malaysia demonstrates the effectiveness and superiority of the B&C algorithm in solving the GDR-GSC model. Numerical experiments show that the customized B&C algorithm can improve the average solving time by 18% while maintaining the same solution quality. Based on realistic cases, we know that horizontal cooperation can increase profits by at least 6.25%.}, archive = {J_COR}, author = {Zheng Wang and Pingyuan Dong and Ying Liu}, doi = {10.1016/j.cor.2025.107293}, journal = {Computers & Operations Research}, month = {1}, pages = {107293}, shortjournal = {Comput. Oper. Res.}, title = {Optimizing production–transportation–delivery in global supply chain with demand ambiguity by branch-and-cut algorithm}, volume = {185}, year = {2026}, }
• (2026). Dynamic crowdsourcing problem in urban–rural distribution using the learning-based approach. COR, 185, 107292. (www)

  @article{ , abstract = {Inspired by real-world urban and rural distribution logistics scenarios, this study explores the dynamic crowdsourcing multi-depot pickup and delivery problem (DCMDPDP) through an online platform (OCP), where requests and crowdsourced vehicles arrive dynamically. Vehicles either collect from multiple depots for deliveries or pick up from customers to depots. To maximize the OCP’s daily total gain, the net value of completed task revenue minus vehicle compensation costs, we integrate anticipated future gains into each decision-making process and formulate the DCMDPDP as a Markov decision process. A learning-based hybrid heuristic algorithm is proposed for the DCMDPDP. Specifically, we develop an enhanced adaptive large neighborhood search algorithm leveraging the heat map to batch orders into multiple groups and assign them to depots, where the heat map is learned offline using a graph convolutional residual network with an attention mechanism model. A value learning-based algorithm is also developed to obtain optimal matches between order batches and vehicles, and near-optimal travel routes. Experimental results demonstrate that the proposed algorithm improves the OCP total gain by 46.09%, 57.13%, 0.49%, 2.45%, 1.08%, and 2.77% over six benchmarks. Furthermore, the proposed algorithm reduces unserved customers to 7.83 on average, outperforming six benchmarks by 2.19–167.52 fewer cases. Moreover, extensive experiments validate that the proposed algorithm is strongly generalizable in handling instances with varying customer sizes and different temporal, spatial, and demand distributions.}, archive = {J_COR}, author = {Zongcheng Zhang and Maoliang Ran and Yanru Chen and M.I.M. Wahab and Mujin Gao and Yangsheng Jiang}, doi = {10.1016/j.cor.2025.107292}, journal = {Computers & Operations Research}, month = {1}, pages = {107292}, shortjournal = {Comput. Oper. Res.}, title = {Dynamic crowdsourcing problem in urban–rural distribution using the learning-based approach}, volume = {185}, year = {2026}, }
• (2026). Towards an unsupervised learning scheme for efficiently solving parameterized mixed-integer programs. COR, 185, 107290. (www)

  @article{ , abstract = {Mixed integer programming (MIP) has been widely utilized to tackle a broad range of real-world decision-making problems, while its solution efficiency remains a key challenge. In this paper, we describe a novel unsupervised learning scheme for accelerating the solution of a family of MIP problems. Distinct substantially from existing learning-to-optimize methods, our proposal seeks to train an autoencoder (AE) for binary variables in an unsupervised learning fashion, using data of optimal solutions to historical instances for a parametric family of MIPs. By a deliberate design of AE architecture and exploitation of its statistical implication, we present a simple and straightforward strategy to construct a class of cutting plane constraints from the decoder parameters of an offline-trained AE. These constraints reliably enclose the optimal binary solutions of new problem instances thanks to the representation strength of AE. More importantly, their integration into the primal MIP problem of an unseen instance leads to a tightened MIP, which can be resolved at decision time using off-the-shelf solvers with much higher efficiency. Our method is applied to two benchmark problems: the batch process scheduling problem, formulated as a mixed-integer linear programming (MILP) problem, and the cart–pole system control problem, formulated as a mixed-integer quadratic programming (MIQP) problem. Comprehensive results demonstrate that our approach significantly reduces the computational cost of off-the-shelf MILP solvers while retaining a high solution quality. The codes of this work are open-sourced at https://github.com/qushiyuan/AE4BV .}, archive = {J_COR}, author = {Shiyuan Qu and Fenglian Dong and Zhiwei Wei and Chao Shang}, doi = {10.1016/j.cor.2025.107290}, journal = {Computers & Operations Research}, month = {1}, pages = {107290}, shortjournal = {Comput. Oper. Res.}, title = {Towards an unsupervised learning scheme for efficiently solving parameterized mixed-integer programs}, volume = {185}, year = {2026}, }
• (2026). Optimal placement of electric vehicle slow-charging stations: A continuous facility location problem under uncertainty. COR, 185, 107289. (www)

  @article{ , abstract = {Electric vehicles (EVs) are becoming a key mechanism to reduce emissions in the transportation industry, and hence contribute to the green transition. In this paper, we present a mathematical programming model which determines the optimal placement of EV charging stations such that chargers are placed in the most cost-efficient way possible for all stakeholders, assuming additionally that EV charging demand is inherently stochastic in nature. The model is formulated as a two-stage, continuous location–allocation model in the form of a generalised Weber problem in two dimensions. However, this formulation is non-convex and notoriously difficult to solve. We therefore propose a suitable discretisation procedure to find high quality solutions in suitable time. The discretisation procedure shows strong performance across a variety of computational experiments using randomly generated scenarios, maintaining robustness in terms of the objective value and overall solution quality. A part of this solution procedure was entered into the 15th AIMMS-MOPTA Optimisation Modelling Competition.}, archive = {J_COR}, author = {H.W. Ljósheim and S. Jenkins and K.D. Searle and J.K. Wolff}, doi = {10.1016/j.cor.2025.107289}, journal = {Computers & Operations Research}, month = {1}, pages = {107289}, shortjournal = {Comput. Oper. Res.}, title = {Optimal placement of electric vehicle slow-charging stations: A continuous facility location problem under uncertainty}, volume = {185}, year = {2026}, }
• (2026). A multi-time-window multi-objective hybrid fleet home health care routing optimization problem considering caregiver utilization and compatibility. COR, 185, 107288. (www)

  @article{ , abstract = {Home healthcare plays a key role in the era of an aging population and limited healthcare resources in hospitals, and one of its important tasks is to develop optimal caregiver visit routines. In the context of sustainable development and dual carbon goals, home healthcare must also address carbon emissions during caregiver visits. One of the most effective initiatives is the use of electric vehicles to gradually replace fuel vehicles as an important means of transportation. Therefore, this paper investigates a multi-objective home healthcare path optimization problem based on a mixed fleet of vehicles. In this problem, the transportation of caregivers is either electric or fuel vehicles, and this paper considers the case of patients with different numbers of multiple time windows, defines compatibility, and optimizes five conflicting objectives under the constraints of time windows, doctor-patient skill matching, electric vehicle battery capacity, compatibility, and maximum working hours. To address this problem, we develop a mixed integer programming model to optimize five objectives: cost minimization, caregiver utilization maximization, workload deviation minimization, patient-caregiver compatibility maximization, and skill level deviation minimization. In addition, this paper proposes a hybrid algorithmic solution model with hybrid simulated annealing and a third-generation non-dominated sorting genetic algorithm and designs two neighborhood structures based on the problem characteristics as well as heuristics for charging station insertion. The results show that the improved hybrid algorithm solves the problem more comprehensively and effectively and can cover a wider solution space with good distribution and diversity.}, archive = {J_COR}, author = {Wendi Li and Gang Du and Xiaohang Yue}, doi = {10.1016/j.cor.2025.107288}, journal = {Computers & Operations Research}, month = {1}, pages = {107288}, shortjournal = {Comput. Oper. Res.}, title = {A multi-time-window multi-objective hybrid fleet home health care routing optimization problem considering caregiver utilization and compatibility}, volume = {185}, year = {2026}, }
• (2026). Modeling and algorithm for job shop scheduling with batch operations in semiconductor fabs. COR, 185, 107287. (www)

  @article{ , abstract = {Semiconductor manufacturing presents a highly complex Job Shop Scheduling Problem (JSP) due to the diversity and large number of processing machines, as well as the intricate manufacturing processes including batch and non-batch operations. Existing studies often either overlook batching problems or address them in oversimplified ways, failing to provide effective solutions for large-scale scheduling challenges with batch operations. For this problem, a model for the JSP involving both batching and non-batching processes in semiconductor fabs is first developed. Then, the First Come First Served (FCFS) approach, as an effective rule-based method, is employed to generate high-quality initial solutions. A tailored Constrained Genetic Algorithm (CGA) by embedding constraints to the stages of genetic algorithms is proposed to further optimize the solution. The CGA incorporates batch grouping, constrained encoding, constrained crossover and constrained mutation to effectively handle the sequential constraints of batch and non-batch processes, ensuring the generation of valid solutions. The CGA is validated using the SMT2020 and SMAT2022 datasets across various scales and scenarios. Experimental results demonstrate that the CGA outperforms FCFS, backward simulation and reinforcement learning. These results highlight the CGA’s effectiveness and robustness in solving complex scheduling problems in semiconductor manufacturing.}, archive = {J_COR}, author = {Wen Ma and Gedong Jiang and Nuogang Sun and Chaoqing Min and Xuesong Mei}, doi = {10.1016/j.cor.2025.107287}, journal = {Computers & Operations Research}, month = {1}, pages = {107287}, shortjournal = {Comput. Oper. Res.}, title = {Modeling and algorithm for job shop scheduling with batch operations in semiconductor fabs}, volume = {185}, year = {2026}, }
• (2026). A bi-criterion steiner traveling salesperson problem with time windows for last-mile electric vehicle logistics. COR, 185, 107286. (www)

  @article{ , abstract = {This paper addresses the problem of energy-efficient and safe routing of last-mile electric freight vehicles. With the rising environmental footprint of the transportation sector and the growing popularity of E-Commerce, freight companies are likely to benefit from optimal time window feasible tours that minimize energy usage while reducing traffic conflicts at intersections and thereby improving safety. We formulate this problem as a Bi-criterion Steiner Traveling Salesperson Problem with Time Windows (BSTSPTW) with energy consumed and the number of left turns at intersections as the two objectives while also considering regenerative braking capabilities. We first discuss an exact mixed-integer programming model with scalarization to enumerate points on the efficiency frontier for small instances. For larger networks, we develop an efficient local search-based heuristic, which uses several operators to intensify and diversify the search process. We demonstrate the utility of the proposed methods using benchmark data and real-world instances from Amazon delivery routes in Austin, US. Comparisons with state-of-the-art solvers show that our heuristics can generate near-optimal solutions within reasonable time budgets, effectively balancing energy efficiency and safety under practical delivery constraints.}, archive = {J_COR}, author = {Prateek Agarwal and Debojjal Bagchi and Tarun Rambha and Venktesh Pandey}, doi = {10.1016/j.cor.2025.107286}, journal = {Computers & Operations Research}, month = {1}, pages = {107286}, shortjournal = {Comput. Oper. Res.}, title = {A bi-criterion steiner traveling salesperson problem with time windows for last-mile electric vehicle logistics}, volume = {185}, year = {2026}, }
• (2026). Maximum capture location problem with random utilities and overflow penalties. COR, 185, 107285. (www)

  @article{ , abstract = {This paper extends the maximum capture location problem with random utilities by incorporating the facility capacity and introducing penalties for overflows into the objective function. We propose a method that combines the key features of two state-of-the-art approaches for the uncapacitated case, which are adapted to solve the problem at hand. The first approach is a linear reformulation that extends the best-known linearization in the literature, which is based on variable substitution. The second approach is a reformulation that incorporates outer-approximation cuts and enhanced submodular cuts, solving the problem via a branch-and-cut approach. We tested the performance of the three approaches on several instances and show that the combined method outperforms each of the preceding techniques. The optimal location patterns of the model are also analysed, and it is found that considering the overflow and overflow penalties in the objective function affects the location decisions. The resulting optimal locations align more closely with practical scenarios.}, archive = {J_COR}, author = {Gonzalo Méndez-Vogel and Sebastián Dávila-Gálvez and Pedro Jara-Moroni and Jorge Zamorano and Vladimir Marianov}, doi = {10.1016/j.cor.2025.107285}, journal = {Computers & Operations Research}, month = {1}, pages = {107285}, shortjournal = {Comput. Oper. Res.}, title = {Maximum capture location problem with random utilities and overflow penalties}, volume = {185}, year = {2026}, }
• (2026). A novel mathematical model for the scheduling of a zero inventory production: An application of process scheduling in fog computing. COR, 185, 107284. (www)

  @article{ , abstract = {One of the main production-related costs in manufacturing is inventory cost since manufacturing firms allocate a vast area to raw material, semi-processed, and final products in production lines and warehouses. Reducing the volume of these inventories leads to lower production-related costs. This paper presents a novel mathematical model for zero-inventory production scheduling. In this model, the jobs arrive at fixed times and are scheduled on a set of unrelated machines. The jobs have different operations that need to be processed one by one. Since the system has zero inventory, the jobs must be processed immediately upon arrival. Also, whenever a job’s operation is complete, the following operation must instantly start (no wait time). That operation is outsourced if no machines are available to process any of the job’s operations. The jobs’ operations are dispatched to the machines from a dispatching center, and there is a latency between the dispatching center, the machines, and the outsourcing center. We present a mixed-integer non-linear programming (MINLP) model to formulate this problem. Then, the MINLP model is turned into a mixed-integer linear programming (MILP) model by linearizing its constraints. Since many production scheduling problems are known to be NP-hard, particularly those involving unrelated parallel machines, precedence constraints, and time-dependent decisions like ours, we adopt two metaheuristics to solve the problem for large-scale cases where exact methods are computationally inefficient. The first is a Genetic Algorithm (GA), and the second is a Teaching-Learning-Based Optimization (TLBO) algorithm. The performance of these algorithms is tested against the optimal solutions obtained from CPLEX for a set of small-scale problems. We consider a real case study, an image processing system, to validate the proposed developments (the MILP model and the GA). The results show that the presented model and algorithm can reduce the system’s total cost by about 12.57% compared to the existing online dispatching rules.}, archive = {J_COR}, author = {Mani Sharifi and Sharareh Taghipour and Abdolreza Abhari and Maciej Rysz}, doi = {10.1016/j.cor.2025.107284}, journal = {Computers & Operations Research}, month = {1}, pages = {107284}, shortjournal = {Comput. Oper. Res.}, title = {A novel mathematical model for the scheduling of a zero inventory production: An application of process scheduling in fog computing}, volume = {185}, year = {2026}, }
• (2026). A unified approach to extract interpretable rules from tree ensembles via integer programming. COR, 185, 107283. (www)

  @article{ , abstract = {Tree ensembles are widely used machine learning models, known for their effectiveness in supervised classification and regression tasks. Their performance derives from aggregating predictions of multiple decision trees, which are renowned for their interpretability properties. However, tree ensemble models do not reliably exhibit interpretable output. Our work aims to extract an optimized list of rules from a trained tree ensemble, providing the user with a condensed, interpretable model that retains most of the predictive power of the full model. Our approach consists of solving a set partitioning problem formulated through Integer Programming. The extracted list of rules is unweighted and defines a partition of the training data, assigning each instance to exactly one rule, and thereby simplifying the explanation process. The proposed method works with tabular or time series data, for both classification and regression tasks, and its flexible formulation can include any arbitrary loss or regularization functions. Our computational experiments offer statistically significant evidence that our method performs comparably to several rule extraction methods in terms of predictive performance and fidelity towards the tree ensemble. Moreover, we empirically show that the proposed method effectively extracts interpretable rules from tree ensembles that are designed for time series data.}, archive = {J_COR}, author = {Lorenzo Bonasera and Emilio Carrizosa}, doi = {10.1016/j.cor.2025.107283}, journal = {Computers & Operations Research}, month = {1}, pages = {107283}, shortjournal = {Comput. Oper. Res.}, title = {A unified approach to extract interpretable rules from tree ensembles via integer programming}, volume = {185}, year = {2026}, }
• (2026). Optimizing high-tech product take-back schemes in a closed-loop supply chain. COR, 185, 107282. (www)

  @article{ , abstract = {Frequent product development is a solution to the shortened product lifecycles in the consumer electronics industry. It enables companies to maintain competitiveness and strengthen their market share. However, environmental concerns bring reverse logistics practices into focus. A take-back policy is a strategic reverse logistics activity known to foster market share; however, it poses various challenges and uncertainties. Considering uncertain demand, we introduced an innovative adoption model with two distinct take-back policies, trade-in and credit, to address challenges in multi-generation production planning. Inspired by real-world practices of companies like Apple and Samsung, our model first examines how trade-in programs drive repeat purchases and enhance market share, with credit-based programs to attract new customers. It then captures changes in demand, production planning, recovery decisions, and internal competition among multiple product generations. Distinct from previous conclusions, this study explores how producers can strategically manage demand for new generations to slow diffusion, thereby increasing refurbishment and recycling volumes over time. Our findings highlight the pivotal role of adaptive pricing strategies and production scalability in maximizing profitability and promoting sustainability in competitive high-tech industries.}, archive = {J_COR}, author = {Fatemeh Keshavarz-Ghorbani and Mohamad Y. Jaber and Seyed Hamid Reza Pasandideh}, doi = {10.1016/j.cor.2025.107282}, journal = {Computers & Operations Research}, month = {1}, pages = {107282}, shortjournal = {Comput. Oper. Res.}, title = {Optimizing high-tech product take-back schemes in a closed-loop supply chain}, volume = {185}, year = {2026}, }
• (2026). Minimizing the weighted number of tardy jobs: Data-driven heuristic for single-machine scheduling. COR, 185, 107281. (www)

  @article{ , abstract = {Existing research on single-machine scheduling is largely focused on exact algorithms, which perform well on typical instances but can significantly deteriorate on certain regions of the problem space. In contrast, data-driven approaches provide strong and scalable performance when tailored to the structure of specific datasets. Leveraging this idea, we focus on a single-machine scheduling problem where each job is defined by its weight, duration, due date, and deadline, aiming to minimize the total weight of tardy jobs. We introduce a novel data-driven scheduling heuristic that combines machine learning with problem-specific characteristics, ensuring feasible solutions, which is a common challenge for ML-based algorithms. Experimental results demonstrate that our approach significantly outperforms the state-of-the-art in terms of optimality gap, number of optimal solutions, and adaptability across varied data scenarios, highlighting its flexibility for practical applications. In addition, we conduct a systematic exploration of ML models, addressing a common gap in similar studies by offering a detailed model selection process and demonstrating why the chosen model is the best fit.}, archive = {J_COR}, author = {Nikolai Antonov and Přemysl Šůcha and Mikoláš Janota and Jan Hůla}, doi = {10.1016/j.cor.2025.107281}, journal = {Computers & Operations Research}, month = {1}, pages = {107281}, shortjournal = {Comput. Oper. Res.}, title = {Minimizing the weighted number of tardy jobs: Data-driven heuristic for single-machine scheduling}, volume = {185}, year = {2026}, }
• (2026). On the periodic service scheduling problem with non-uniform demands. COR, 185, 107280. (www)

  @article{ , abstract = {This paper introduces the Periodic Service Scheduling Problem with Non-uniform Demands, in which the best service policy for a set of customers with periodically recurring demand through a given finite planning horizon has to be determined. Service to customers is provided at every time period by a set of potential service providers, each of them with an activation cost and a capacity. The decisions to be made include the servers to be activated at each time period together with a service schedule and server allocation for every customer that respect the periodicity of customer demand and the capacity of the activated servers, which minimize the total cost of the activated servers. We give a first Integer Linear Programming formulation with one set of decision variables associated with each of the decisions of the problem. Afterwards, we develop a logic-based Benders reformulation where one set of variables is projected out and constraints that guarantee the feasibility of the solutions are introduced. The separation problem for the new set of constraints is studied, and an exact Branch & Logic-Benders-Cut algorithm for the reformulation is proposed together with several variations and enhancements. The particular cases in which all servers are identical and in which all parameters are time-invariant are also studied. Extensive computational experiments assess the superiority of the logic-based Benders reformulation over the first formulation.}, archive = {J_COR}, author = {Elena Fernández and Jörg Kalcsics}, doi = {10.1016/j.cor.2025.107280}, journal = {Computers & Operations Research}, month = {1}, pages = {107280}, shortjournal = {Comput. Oper. Res.}, title = {On the periodic service scheduling problem with non-uniform demands}, volume = {185}, year = {2026}, }
• (2026). Q-learning-based hyper-heuristic algorithm for open dimension irregular packing problems. COR, 185, 107279. (www)

  @article{ , abstract = {Heuristic methods provide a computationally efficient framework for addressing two-dimensional irregular packing problems, particularly in resource-constrained industrial settings. As a typical combinatorial optimization problem, irregular packing exhibits exponential growth in computational complexity with increasing workpiece counts, while the solution space dynamically reconfigures due to geometric variability among workpieces. Although heuristic algorithms can generate feasible layouts within acceptable timeframes, their reliance on fixed search rule limits adaptability to diverse scenarios, necessitating more flexible approaches. In this paper, a hyper-heuristic algorithm based on Q-Learning is proposed to solve open dimension packing problems, including one-open and two-open dimension problems. Q-Learning is adopted as the high-level strategy for its ability to optimize low-level heuristic selection through reward-driven experience accumulation. The method incorporates a mixed encoding method for solution representation, four specialized low-level heuristic operators, a linear population decline mechanism, and an elite preservation strategy to balance exploration–exploitation. The Q-Learning controller dynamically selects operators by updating the Q-table based on Bellman’s equation. The proposed algorithm is compared to some advanced algorithms in general datasets. The results show that our method has better performance and applicability.}, archive = {J_COR}, author = {Yongchun Wang and Qingjin Peng and Zhen Wang and Shuiquan Huang and Zhengkai Xu and Chuanzhen Huang and Baosu Guo}, doi = {10.1016/j.cor.2025.107279}, journal = {Computers & Operations Research}, month = {1}, pages = {107279}, shortjournal = {Comput. Oper. Res.}, title = {Q-learning-based hyper-heuristic algorithm for open dimension irregular packing problems}, volume = {185}, year = {2026}, }
• (2026). Hybrid modelling using simulation and machine learning in healthcare. COR, 185, 107278. (www)

  @article{ , abstract = {Modelling & Simulation (M&S) and Machine Learning (ML) methodologies have undergone significant advancements, enabling transformative applications across various industries. The integration of M&S and ML into a Hybrid M&S-ML approach leverages the unique strengths of both fields, offering enhanced model precision, improved efficiency, and more effective decision support. This review explores the increasing convergence of ML algorithms with traditional M&S methods- namely Agent-Based Modelling & Simulation, Discrete Event Simulation, and System Dynamics- in healthcare applications. Through a systematic review of 90 relevant studies, this article provides a comprehensive synthesis of the current state-of-the-art Hybrid M&S-ML in healthcare. Specifically, it examines the M&S and ML methodologies employed, associated software tools and programming languages, analyses integration patterns and data exchange mechanisms, and explores application domains, as well as the types and motivations for hybridisation. Key findings highlight prominent methodological and technical trends, as well as opportunities for combining M&S with ML to address healthcare challenges. These insights provide direction for modellers and data scientists in developing hybrid M&S–ML approaches that more effectively combine simulation capabilities with data-driven learning. The review also demonstrates the potential of such approaches to advance methodological innovation and support evidence-based decision-making in diverse healthcare contexts.}, archive = {J_COR}, author = {Ali Ahmadi and Masoud Fakhimi and Carin Magnusson}, doi = {10.1016/j.cor.2025.107278}, journal = {Computers & Operations Research}, month = {1}, pages = {107278}, shortjournal = {Comput. Oper. Res.}, title = {Hybrid modelling using simulation and machine learning in healthcare}, volume = {185}, year = {2026}, }
• (2026). A note on battery swapping policies in the electric vehicle routing problem with time windows and battery swapping vehicles. COR, 185, 107277. (www)

  @article{ , abstract = {Çatay and Sadati [An improved matheuristic for solving the electric vehicle routing problem with time windows and synchronized mobile charging/battery swapping. Computers & Operations Research 159, 106310, 2023] explores a variant of the Electric Vehicle Routing Problem with Time Windows that incorporates mobile chargers for recharging electric vehicles (EVs) at selected locations while serving customers. The authors propose a matheuristic method to address this problem and its special case, where EV batteries are swapped in constant time instead of being recharged over variable durations. While comparing their results with those in the literature, the authors overlook a critical assumption regarding the swapping policy, potentially causing confusion in interpreting the findings. This note addresses the issue, clarifies the overlooked assumption, and updates the results that do not align with the actual scenario in the literature. Furthermore, it introduces two new battery swapping policies and presents an extensive computational study to offer new insights on synchronized mobile battery swapping.}, archive = {J_COR}, author = {Bülent Çatay and İhsan Sadati}, doi = {10.1016/j.cor.2025.107277}, journal = {Computers & Operations Research}, month = {1}, pages = {107277}, shortjournal = {Comput. Oper. Res.}, title = {A note on battery swapping policies in the electric vehicle routing problem with time windows and battery swapping vehicles}, volume = {185}, year = {2026}, }
• (2026). Solving the strip packing problem with a decomposition framework and a generic solver: Implementation, tuning, and reinforcement-learning-based hybridization. COR, 185, 107276. (www)

  @article{ , abstract = {In the strip packing problem, the objective is to pack a set of two-dimensional items into a strip of fixed width such that the total height of the packing is minimized. The current state-of-the-art exact approach for the problem uses a decomposition framework in which the main problem (MP) fixes the item abscissas and the strip height, whereas the subproblem (SP) determines whether a set of item ordinates resulting in a feasible packing exists. Even though this decomposition framework has already been used several times in the literature, implementation details were often obfuscated, limiting the outreach of the approach. We address this issue by thoroughly describing and testing various builds for this framework, investigating important features such as the way to forbid an infeasible solution in the MP (e.g., by rejecting them or through a no-good cut) and the techniques used to solve the MP and the SP. One of our findings is that a minor implementation tweak such as changing the random seed between two MP iterations can bring the same level of improvement as a more involved feature such as strengthening the no-good cuts. From our extensive experiments, we identify two versions of the framework that produce complementary results: one where the main problem is solved with integer linear programming and the other where it is solved with constraint programming. We then train a reinforcement learning agent to find the best hybridization of these two algorithms and show that the resulting approach obtains state-of-the-art results on benchmark instances.}, archive = {J_COR}, author = {Fatih Burak Akçay and Maxence Delorme}, doi = {10.1016/j.cor.2025.107276}, journal = {Computers & Operations Research}, month = {1}, pages = {107276}, shortjournal = {Comput. Oper. Res.}, title = {Solving the strip packing problem with a decomposition framework and a generic solver: Implementation, tuning, and reinforcement-learning-based hybridization}, volume = {185}, year = {2026}, }
• (2026). An adaptive K-means and reinforcement learning (RL) algorithm to effective vaccine distribution. COR, 185, 107275. (www)

  @article{ , abstract = {We present a new adaptive reinforcement learning (RL) approach, integrated with a K-means clustering algorithm and guided by simulated annealing, to address the capacitated vehicle routing for vaccine distribution (CVRVD) problem. This integrated method provides an efficient and scalable solution for optimizing vaccine distribution logistics. By incorporating cost factors related to travel distance, inventory levels, and penalty terms – while adhering to delivery time windows – our approach improves both operational efficiency and vaccine allocation effectiveness. Experimental results demonstrate that our K-means supported RL algorithm significantly outperforms traditional solvers in tackling this NP-hard problem, particularly in large-scale scenarios. Specifically, our approach can efficiently solve CVRVD instances with up to 1,000 facilities—scenarios that are computationally intractable for exact methods. We demonstrate the effectiveness of the adaptive K-means supported RL algorithm using data from New Jersey, USA, where facility-level vaccination data were available through the state’s Immunization Information System. Beyond vaccine distribution, our method has broad applicability in logistics and transportation, enabling more efficient and cost-effective allocation of critical resources such as vaccines and medical supplies.}, archive = {J_COR}, author = {Elson Cibaku and İ. Esra Büyüktahtakın}, doi = {10.1016/j.cor.2025.107275}, journal = {Computers & Operations Research}, month = {1}, pages = {107275}, shortjournal = {Comput. Oper. Res.}, title = {An adaptive K-means and reinforcement learning (RL) algorithm to effective vaccine distribution}, volume = {185}, year = {2026}, }
• (2026). A Q-learning-based multi-objective hyper-heuristic algorithm for energy-efficient integrated distributed hybrid flow-shop scheduling with preventive maintenance. COR, 185, 107267. (www)

  @article{ , abstract = {Driven by the dual engines of supply chain integration and low-carbon transformation in industrial Internet of Things (IIoT) manufacturing systems, energy-efficient integrated distributed scheduling has emerged as a pivotal component of industrial intelligence-driven smart manufacturing. This article investigates the energy-efficient integrated distributed hybrid flow shop scheduling problem with preventive maintenance (EE-IDHFSP-PM), which aims to minimize the dual objectives of makespan and total carbon emissions. In this study, a mixed-integer linear programming (MILP) model is established for the EE-IDHFSP-PM, making the first attempt to solve such NP-hard problem by using a Q -learning-based multi-objective hyper-heuristic algorithm (QLMHHA). First, a modified NEH-based initialization method is introduced to produce high-quality solutions that balance multiple optimization objectives, ensuring both the quality and diversity of initial populations. Second, a novel multi-stage collaborative energy-efficient strategy (MSC_EES) is developed to dynamically adjust the processing speeds of machines on non-critical paths, which reduces energy consumption across stages. Third, a new Q -learning-based high-level strategy (HLS) is devised to dynamically coordinate twelve low-level heuristics (LLHs) according to specific states, improving adaptive search efficiency through superior exploration–exploitation trade-offs. Fourth, a dual-criterion reward mechanism is proposed to evaluate population quality in terms of both convergence and diversity, which can deliver immediate feedback and effectively guide evolutionary processes. Fifth, comprehensive convergence and computational complexity analyses of critical components are conducted to confirm the stability, reliability, and efficiency of QLMHHA. Extensive experiments are carried out on 54 small-scale and 24 large-scale instances, which demonstrate that QLMHHA achieves promising performance in both effectiveness and efficacy against state-of-the-art multi-objective algorithms for addressing the EE-IDHFSP-PM. These findings validate the efficacy and superiority of QLMHHA in tackling complex scheduling challenges, providing valuable theoretical implications and practical insights for energy-efficient distributed manufacturing systems.}, archive = {J_COR}, author = {Zi-Qi Zhang and Xin-Yun Wu and Bin Qian and Rong Hu and Jian-Bo Yang}, doi = {10.1016/j.cor.2025.107267}, journal = {Computers & Operations Research}, month = {1}, pages = {107267}, shortjournal = {Comput. Oper. Res.}, title = {A Q-learning-based multi-objective hyper-heuristic algorithm for energy-efficient integrated distributed hybrid flow-shop scheduling with preventive maintenance}, volume = {185}, year = {2026}, }
• (2026). A Q-learning-based evolutionary algorithm for solving the low-carbon multi-objective flexible job shop scheduling problem. COR, 185, 107266. (www)

  @article{ , abstract = {In recent years, how to reduce energy consumption at the manufacturing system level in the low-carbon multi-objective flexible job shop scheduling problem (LCM-FJSP) has received significant attention. In this research, a model with the maximum completion time, total machine workload and total carbon emissions is built. Moreover, a Q-learning-based adaptive weight-adjusted decomposition evolutionary algorithm (QMOEA/D-AWA) is proposed. In the QMOEA/D-AWA, an initialization strategy with four heuristic initial rules for obtaining high-quality population, a variable neighborhood search strategy with four problem-specific local search methods for enhancing exploration and a Q-learning-based parameter adaptive strategy for automatically determining the number of neighborhood solutions are designed. To validate the effectiveness of the proposed QMOEA/D-AWA, it is compared with five state-of-the-art algorithms on 15 instances. In the statistical analysis, the QMOEA/D-AWA obtains the overwhelming metric results in 10 instances. In the visual analysis, the completion time is reduced by 3.74%, the total workload is reduced by 3.94%, and the carbon emissions are reduced by 5.94%.}, archive = {J_COR}, author = {Zhixue Wang and Maowei He and Hanning Chen and Yabao Hu and Yelin Xia}, doi = {10.1016/j.cor.2025.107266}, journal = {Computers & Operations Research}, month = {1}, pages = {107266}, shortjournal = {Comput. Oper. Res.}, title = {A Q-learning-based evolutionary algorithm for solving the low-carbon multi-objective flexible job shop scheduling problem}, volume = {185}, year = {2026}, }
• (2026). A benders-branch-and-cut methodology for global cargo vessel traffic prediction given declining arctic sea ice and changing risks. COR, 185, 107265. (www)

  @article{ , abstract = {Global warming has led to declining sea-ice in the Arctic Ocean, making it easier for ice-class vessels to navigate Arctic waters for greater portions of the year. As sailing conditions in these waters improve over coming decades, these passageways are expected to open for larger portions of the year and to become increasingly viable options for unsupported transit and even open-water vessels. This paper proposes a Benders-branch-and-cut methodology for estimating changes in global maritime cargo flow patterns under future climate scenarios with declining Arctic sea ice. The model accounts for changing incident risk along Arctic passageways and corresponding ice-class vessel and icebreaker escort requirements, lower speeds, increased insurance premiums, higher accident probabilities, and constraints on path-based maximum risk exposure. The resulting mixed-integer program involves path-based, continuous decision variables. The solution technique is applied on a model of the global maritime container network including 80 ports, 76 routes, 426 links and 4,303 legs associated with the world’s largest carrier alliance. Embedded acceleration techniques and a label-correcting algorithm that employs specialized fathoming rules for a non-additive, constrained path subproblem enable solution at this global scale. The outcome is an estimate of seasonal future global maritime trade flows along key global routes and through ports predicted under six climate-related scenarios. Results illustrate that the developed model can provide support to companies, nations and regions as they prepare for a changing global landscape and climate.}, archive = {J_COR}, author = {Wenjie Li and Elise Miller-Hooks}, doi = {10.1016/j.cor.2025.107265}, journal = {Computers & Operations Research}, month = {1}, pages = {107265}, shortjournal = {Comput. Oper. Res.}, title = {A benders-branch-and-cut methodology for global cargo vessel traffic prediction given declining arctic sea ice and changing risks}, volume = {185}, year = {2026}, }
• (2026). Mitigating choice model ambiguity: A consensus framework and its application to assortment optimization. COR, 185, 107264. (www)

  @article{ , abstract = {Discrete choice models have become a popular tool to accurately predict complex choice behavior. Due to a variety of possible error sources, estimated choice models tend to be subject to ambiguity, inducing different optimal decisions of highly varying quality. This study aims at mitigating choice model ambiguity associated with a given set of models in terms of their ability to yield optimal decisions. We propose a framework and a set of performance metrics to assess the reliability of choice models and their induced decisions. The use of this framework is then exemplified in the context of rank-based choice models for assortment optimization. Extensive sets of numerical results suggest that our proposed approaches indeed allow decision-makers to identify choice models that are likely to produce high quality decisions, boosting confidence in using choice models in practice. While robust optimization on the original set of choice models tends to be rather conservative, we then use the proposed metrics to reduce the size of the ambiguity set, allowing us to improve the expected assortment quality and the overall downside risk. Given the practical usefulness of robust optimization in this context, we further propose a decomposition algorithm, solving the optimization problem in a fraction of the original time and revealing that only a few among a large set of choice models are determinant in optimal robust solutions.}, archive = {J_COR}, author = {Öykü Naz Attila and Sanjay Dominik Jena and Walter Rei}, doi = {10.1016/j.cor.2025.107264}, journal = {Computers & Operations Research}, month = {1}, pages = {107264}, shortjournal = {Comput. Oper. Res.}, title = {Mitigating choice model ambiguity: A consensus framework and its application to assortment optimization}, volume = {185}, year = {2026}, }
• (2026). A deep reinforcement learning approach for dynamic job-shop scheduling problem considering time variable and new job arrivals. COR, 185, 107263. (www)

  @article{ , abstract = {In recent years, the complexity of the production process due to increased demand for customization has greatly increased the difficulty of dynamic job-shop scheduling problem (DJSP). This paper proposes a deep reinforcement learning (DRL) approach to tackle the DJSP based on proximal policy optimization (PPO) algorithm. A novel state representation method that expresses state features as multi-channel images is proposed to simplify the state characterization process. Various heuristic-based priority dispatching rules (PDRs)are used to construct action space. By converting scheduling instances into images and leveraging the spatial pyramid pooling fast (SPPF) module for feature extraction, this model can handle scheduling instances of varying scales and map size-independent processing information matrix to fixed action space. Additionally, a dense reward based on a predefined scheduling region is developed to offer detailed guidance to the agent, enabling more precise and comprehensive policy assessment. Static tests are conducted on well-known benchmarks, and the experimental results indicate that our scheduling model surpasses the performance of the three latest DRL approaches on average. Compared with PDR methods, dynamic experiments demonstrate that the proposed DRL model excels in adaptability and robustness when new tasks arrive and the processing time fluctuates with uncertainty.}, archive = {J_COR}, author = {Haoyang Yu and Wenbin Gu and Na Tang and Zhenyang Guo}, doi = {10.1016/j.cor.2025.107263}, journal = {Computers & Operations Research}, month = {1}, pages = {107263}, shortjournal = {Comput. Oper. Res.}, title = {A deep reinforcement learning approach for dynamic job-shop scheduling problem considering time variable and new job arrivals}, volume = {185}, year = {2026}, }
• (2026). An efficient resource utilization and scheduling strategy for in-service aircraft maintenance and operations. COR, 185, 107262. (www)

  @article{ , abstract = {Optimal scheduling of maintenance activities requires the solution of combinatorial optimization problems that need to be efficiently modeled and solved with optimization techniques. Maintenance scheduling and operations-associated problems in the aviation industry can efficiently enhance competitiveness. In the maintenance and scheduling problem, aircrafts need to undergo tasks for both line (A check) and base (C check) maintenance at various hangers at MRO (Maintenance, Repair and Operations) based on resource availability (both human and material). The determination of the optimal maintenance plan, in terms of allocating the resources to the aircraft, and resource movement from one aircraft to another based on availability and licensed skills in the presence of multiple tasks and capacity constraints so as to obtain maximum utilization of resources at maintenance site and minimize the turnaround time is a complex combinatorial optimization problem. To the best of our knowledge, this work is the first CP (Constraint Programming) based mathematical solution that jointly integrates zone, task precedence, technician-pool sharing, and multi-shift continuity for large-scale aircraft maintenance scheduling. In this article, we proposed an efficient optimization strategy that overcomes many of the drawbacks of the formulation/strategies available in literature and helps in determining efficient execution of maintenance work packages. The proposed strategy is generic, encompassing multi-aircraft, multi-skill and multi-shift scheduling capabilities and is validated on two real scenario business case studies, one each for line maintenance (A check) tasks and base maintenance (C check) tasks, as well as six large-scale synthetic scenarios with up to 20,000 tasks, demonstrating feasibility and scalable performance. The proposed strategy is demonstrated on MRO scheduling and it shows an improvement of up to 30.68% in the turn-around time by incorporating the proposed optimization strategy.}, archive = {J_COR}, author = {Sandeep Singh Chauhan and Likhith Maadhav and Abhijit Dake and Gauthier Brillaud}, doi = {10.1016/j.cor.2025.107262}, journal = {Computers & Operations Research}, month = {1}, pages = {107262}, shortjournal = {Comput. Oper. Res.}, title = {An efficient resource utilization and scheduling strategy for in-service aircraft maintenance and operations}, volume = {185}, year = {2026}, }
• (2026). A data-driven optimization approach for the integrated train scheduling and maintenance planning in high-speed railways. COR, 185, 107261. (www)

  @article{ , abstract = {In railway systems, preventive maintenance plans are essential for ensuring the safety of train operations. However, these tasks are often subject to various disturbances (e.g., bad weather), leading to unpredictable deviations between planned and actual maintenance durations, which can further disrupt train schedules. Unlike most studies that assume constant maintenance durations, this paper introduces a data-driven, two-stage distributionally robust optimization (DRO) model for jointly optimizing train scheduling and maintenance planning. In the first stage, we determine the initial train schedule and maintenance plan. In the second stage, we allow for slight adjustments to train departure and arrival times at each station to accommodate disturbances affecting maintenance tasks. Our objective is to minimize both the expected travel time of trains and the deviation from the planned schedule under worst-case scenarios for maintenance disturbances. To capture the uncertainty of maintenance disturbances, we construct an ambiguity set using historical data and the Wasserstein metric. We show that the proposed two-stage DRO model, formulated over the Wasserstein ambiguity set, can be reformulated into an efficiently solvable equivalent form. Finally, we apply our model to a real-world case study of the Beijing–Guangzhou high-speed railway and compare it with traditional stochastic programming methods, including sample average approximation and robust optimization. The results highlight the efficiency of our approach and provide valuable insights for railway management.}, archive = {J_COR}, author = {Hangyu Ji and Chuntian Zhang and Jiateng Yin and Lixing Yang}, doi = {10.1016/j.cor.2025.107261}, journal = {Computers & Operations Research}, month = {1}, pages = {107261}, shortjournal = {Comput. Oper. Res.}, title = {A data-driven optimization approach for the integrated train scheduling and maintenance planning in high-speed railways}, volume = {185}, year = {2026}, }
• (2026). ROBIST: Robust optimization by iterative scenario sampling and statistical testing. COR, 185, 107260. (www)

  @article{ , abstract = {In this paper, we propose ROBIST , a simple, yet effective, data-driven algorithm for optimization under parametric uncertainty. The algorithm first generates solutions in an iterative manner by sampling and optimizing over a relatively small set of scenarios. Then, using statistical testing, the robustness of the solutions is evaluated, which can be done with a much larger set of scenarios. ROBIST offers a number of practical advantages over existing methods as it is: (i) easy to implement, (ii) able to deal with a wide range of problems and (iii) capable of providing sharp probability guarantees that are easily computable and independent of the dimensions of the problem. Numerical experiments demonstrate the effectiveness of ROBIST in comparison to alternative methods.}, archive = {J_COR}, author = {Justin Starreveld and Guanyu Jin and Dick den Hertog and Roger J.A. Laeven}, doi = {10.1016/j.cor.2025.107260}, journal = {Computers & Operations Research}, month = {1}, pages = {107260}, shortjournal = {Comput. Oper. Res.}, title = {ROBIST: Robust optimization by iterative scenario sampling and statistical testing}, volume = {185}, year = {2026}, }
• (2026). A branch-and-price algorithm for energy aware task scheduling of constellations of nanosatellites. COR, 185, 107259. (www)

  @article{ , abstract = {This paper presents a branch-and-price algorithm for solving the Optimal Network Task Scheduling (ONTS) problem in satellite constellations. The algorithm efficiently manages both constellation tasks that can be performed by any satellite and satellite-specific tasks that must be executed by designated satellites, while considering critical energy constraints. We formulate the problem as a Mixed-Integer Linear Programming (MILP) model and develop a Dantzig–Wolfe decomposition that handles battery management constraints for the satellites at the master level, while addressing constellation-wide coordination requirements in the subproblems. A novel dynamic programming algorithm is proposed to solve the pricing subproblem for constellation tasks, augmented with dual stabilization techniques to improve convergence. Comprehensive computational experiments on realistic instances derived from nanosatellite operations demonstrate the effectiveness of the algorithm. Results show that our structured formulation significantly outperforms a naive approach, particularly for large instances, while effectively balancing workload distribution and energy management across the constellation. This work provides a practical framework for optimizing task scheduling in modern satellite constellations, with direct applications in Earth observation, telecommunications, and scientific missions.}, archive = {J_COR}, author = {Pedro Marcolin Antunes and Laio Oriel Seman and Eduardo Camponogara}, doi = {10.1016/j.cor.2025.107259}, journal = {Computers & Operations Research}, month = {1}, pages = {107259}, shortjournal = {Comput. Oper. Res.}, title = {A branch-and-price algorithm for energy aware task scheduling of constellations of nanosatellites}, volume = {185}, year = {2026}, }
• (2026). A branch-and-cut algorithm for the pallet-loading vehicle routing problem considering load balance of semi-trailer trucks. COR, 185, 107258. (www)

  @article{ , abstract = {The work presented in this paper is motivated by some practical applications, introducing and solving a new problem called the split delivery vehicle routing problem with time windows and two-dimensional loading constraints considering load balance (2L-SDVRPTW-LB). This problem expands the integrated routing and loading problem, in which the required pallet orders are delivered to customers by a fleet of homogeneous semi-trailer trucks. Firstly, with the goal to minimize the total transportation cost, a mixed-integer linear programming model for the 2L-SDVRPTW-LB is proposed. Secondly, an exact algorithm framework based on the branch-and-cut algorithm is designed to solve the problem. In the algorithm framework, a metaheuristic algorithm is designed to obtain the upper bound. Subsequently, various classes of valid inequalities are introduced to strengthen the formulation, while the constraints ensuring load balance are applied in a lazy manner to optimize computational efficiency. Numerical experiments based on modified benchmark instances demonstrate that the proposed branch-and-cut algorithm effectively solves the 2L-SDVRPTW-LB and its variant of excluding split delivery. This study provides valuable insights for addressing the pallet-loading vehicle routing problem considering load balance of semi-trailer trucks.}, archive = {J_COR}, author = {Xiangbin Xu and Haoxing Ouyang}, doi = {10.1016/j.cor.2025.107258}, journal = {Computers & Operations Research}, month = {1}, pages = {107258}, shortjournal = {Comput. Oper. Res.}, title = {A branch-and-cut algorithm for the pallet-loading vehicle routing problem considering load balance of semi-trailer trucks}, volume = {185}, year = {2026}, }
• (2026). A hybrid memetic metaheuristic for medical staff assignment in major public health emergencies. COR, 185, 107256. (www)

  @article{ , abstract = {During major public health emergencies, effective assignment of medical staff is crucial for saving lives and controlling the spread of epidemics. This work focuses on the assignment of doctors and nurses to hospitals to form treatment groups that carry out patient treatment tasks. We consider the practical constraints of skill types of medical staff and the severity of patients’ conditions and propose a mixed integer programming model with the objective of maximizing demand satisfaction and personnel skill matching. To solve this problem, we introduce a hybrid memetic search algorithm that combines a specialized crossover operator for generating promising offspring solutions and a variable neighborhood search procedure to improve their quality. Computational results demonstrate that our algorithm outperforms the general mixed integer programming solver GUROBI . The key components of the proposed algorithm are experimentally analyzed and managerial insights are derived.}, archive = {J_COR}, author = {Yang Wang and He Zheng and Zequn Wei and Christophe Wilbaut and Saïd Hanafi}, doi = {10.1016/j.cor.2025.107256}, journal = {Computers & Operations Research}, month = {1}, pages = {107256}, shortjournal = {Comput. Oper. Res.}, title = {A hybrid memetic metaheuristic for medical staff assignment in major public health emergencies}, volume = {185}, year = {2026}, }
• (2026). Learning-guided iterated local search for the minmax multiple traveling salesman problem. COR, 185, 107255. (www)

  @article{ , abstract = {The minmax multiple traveling salesman problem involves minimizing the costs of a longest tour among a set of tours. The problem is of great practical interest because it can be used to formulate several real-life applications. To solve this computationally challenging problem, we propose a learning-driven iterated local search approach that combines an effective local search procedure to find high-quality local optimal solutions and a multi-armed bandit algorithm to select removal and insertion operators to escape local optimal traps. Extensive experiments on 77 commonly used benchmark instances show that the algorithm achieves excellent results in terms of solution quality and running time. In particular, it achieves 32 new best results (improved upper bounds) and matches the best-known results for 35 other instances. Additional experiments shed light on the understanding of the algorithm’s constituent elements. Multi-armed bandit selection can be used advantageously in other multi-operator local search algorithms.}, archive = {J_COR}, author = {Pengfei He and Jin-Kao Hao and Jinhui Xia}, doi = {10.1016/j.cor.2025.107255}, journal = {Computers & Operations Research}, month = {1}, pages = {107255}, shortjournal = {Comput. Oper. Res.}, title = {Learning-guided iterated local search for the minmax multiple traveling salesman problem}, volume = {185}, year = {2026}, }
• (2026). Grouping strategies on two-phase methods for bi-objective combinatorial optimization. COR, 185, 107254. (www)

  @article{ , abstract = {Two-phase methods are commonly used to solve bi-objective combinatorial optimization problems. In the first phase, all extreme supported nondominated points are generated through a dichotomic search. This phase also allows the identification of search zones that may contain other nondominated points. The second phase focuses on exploring these search zones to locate the remaining points, which typically accounts for most of the computational cost. Ranking algorithms are frequently employed to explore each zone individually, but this approach leads to redundancies, causing multiple visits to the same solutions. To mitigate these redundancies, we propose several strategies that group adjacent zones, allowing a single run of the ranking algorithm for the entire group. Additionally, we explore an implicit grouping approach based on a new concept of coverage. Our experiments on the Bi-Objective Spanning Tree Problem demonstrate the beneficial impact of these grouping strategies when combined with coverage.}, archive = {J_COR}, author = {Felipe O. Mota and Luís Paquete and Daniel Vanderpooten}, doi = {10.1016/j.cor.2025.107254}, journal = {Computers & Operations Research}, month = {1}, pages = {107254}, shortjournal = {Comput. Oper. Res.}, title = {Grouping strategies on two-phase methods for bi-objective combinatorial optimization}, volume = {185}, year = {2026}, }
• (2026). Drone-aided mobile blood collection problem: A rolling-horizon-based matheuristic. COR, 185, 107253. (www)

  @article{ , abstract = {This study introduces the drone-aided mobile blood collection problem, which integrates mobile blood donation vehicles with drones to improve operations related to the blood collection in urban areas. Each vehicle, carrying multiple drones, travels to several collection sites to conduct blood collection operations within a working day. Drones fly between vehicles to pick up collected blood bags and deliver them to the blood center. This collaborative framework enhances the performances of the collection system and ensures the freshness of collected blood upon arrival to the blood center. We develop a novel mixed-integer linear programming model to optimally synchronize the routes and collection schedules of mobile units and drones to ensure the timely delivery of collected blood to the blood center. We also develop a rolling-horizon-based matheuristic to solve large-scale instances of the problem. This algorithm combines a rolling horizon approach, which divides the problem into manageable subproblems solved sequentially, with a local branching technique that enhances solutions by exploring promising neighborhoods. To evaluate the algorithm’s performance, we conduct a comprehensive computational study. Our results show that the proposed algorithm not only finds better solutions than those obtained by Gurobi but also outperforms other matheuristics, including the rolling horizon, relax-and-fix, and fix-and-optimize algorithms. Finally, we demonstrate the real-life applicability of the problem through a case study in Quebec City, Canada.}, archive = {J_COR}, author = {Amirhossein Abbaszadeh and Hossein Hashemi Doulabi}, doi = {10.1016/j.cor.2025.107253}, journal = {Computers & Operations Research}, month = {1}, pages = {107253}, shortjournal = {Comput. Oper. Res.}, title = {Drone-aided mobile blood collection problem: A rolling-horizon-based matheuristic}, volume = {185}, year = {2026}, }
• (2026). Fair and efficient multi-agent routing for cooperative and autonomous agricultural fleets with implements. COR, 185, 107252. (www)

  @article{ , abstract = {The growing use of autonomous tractor fleets with detachable implements presents complex logistical challenges in agriculture. Current systems often rely on simple heuristics and avoid implement swapping, limiting efficiency. A central challenge is to dynamically coordinate vehicle routing and implement exchanges to enable efficient, low-intervention task execution. Due to high costs, such fleets are owned mainly by large enterprises or cooperatives, where fair task allocation and profit sharing are critical. Addressing both coordination and fairness, in this paper, we introduce the Agricultural Fleet Vehicle Routing Problem with Implements (AFVRPI). We propose a distributed model derived from a centralized formulation also presented in this paper. This model is embedded within a Distributed Multi-Agent System Architecture (DIMASA), where autonomous vehicle agents manage routing and implement use under limited fuel autonomy, while implement agents ensure compatibility and sufficient capacity to meet task demands. Our solution applies systematic egalitarian social welfare optimization to iteratively maximize the profit of the worst-off vehicle, balancing fairness with system efficiency. To enhance scalability, we use column generation in the distributed model, achieving solution quality comparable to the centralized model while significantly reducing computing time. Simulation results on new benchmark instances demonstrate that our distributed multi-agent AFVRPI approach is scalable, efficient, and fair.}, archive = {J_COR}, author = {Aitor López-Sánchez and Marin Lujak and Frédéric Semet and Holger Billhardt}, doi = {10.1016/j.cor.2025.107252}, journal = {Computers & Operations Research}, month = {1}, pages = {107252}, shortjournal = {Comput. Oper. Res.}, title = {Fair and efficient multi-agent routing for cooperative and autonomous agricultural fleets with implements}, volume = {185}, year = {2026}, }
• (2026). Flexible scheduling of customized bus for green mega-events: A distributionally robust optimization approach. COR, 185, 107249. (www)

  @article{ , abstract = {Mega-events such as the Olympics and the World Championships face significant challenges in evacuating large numbers of attendees after the events conclude, which consume substantial transportation resources. Under the global pressure to reduce carbon emissions, energy conservation and emission reduction are increasingly becoming top priorities. This paper focuses on the efficient scheduling of customized buses (CB) after green mega-events, incorporating skip-stop operations and coordinated bus services to minimize energy consumption, fixed and transportation costs, and facilitate the evacuation of attendees, while accounting for practical constraints such as the availability of customized buses, vehicle capacity, time windows, and flow balance. A distributionally robust optimization (DRO) model is developed, using a novel ambiguity set to model uncertain demand via parametric interval-valued fuzzy variables. To ensure computational tractability, the model is reformulated as an integer linear programming model. To address the computational challenges of large-scale instances, an improved variable neighborhood search heuristic is designed by incorporating the reinforcement learning techniques, including the KL-UCB algorithm and a sliding window mechanism. Extensive numerical experiments are conducted to verify the performance of the proposed heuristic. Computational results demonstrate that the proposed DRO model effectively handles uncertainty, offering robust and adaptable solutions. Compared to existing heuristics, the proposed heuristic improves performance by 6.51% on average, and incorporating reinforcement learning into VNS enhances computational efficiency by 4.88% on average. A real-life case study further validates the model, demonstrating that the skip-stop strategy significantly reduces vehicle travel time and enhances overall operational efficiency.}, archive = {J_COR}, author = {Xiaojie An and Xiang Li and Bowen Zhang}, doi = {10.1016/j.cor.2025.107249}, journal = {Computers & Operations Research}, month = {1}, pages = {107249}, shortjournal = {Comput. Oper. Res.}, title = {Flexible scheduling of customized bus for green mega-events: A distributionally robust optimization approach}, volume = {185}, year = {2026}, }
• (2026). Optimal planning of power distribution networks with fault-tolerant configuration. COR, 185, 107248. (www)

  @article{ , abstract = {Power Distribution networks are essential infrastructures that should be designed by satisfying two conflicting requests: cost minimization and reliability. While traditional network planning aimed at radial configurations, which are more similar to the typical working configuration of a network but are not fault-tolerant, modern techniques seek for meshed configurations, since these architectures are more fault-tolerant. Due to the complexity of the problem and the large size of nowadays instances, most of the techniques used for planning are based on heuristic approaches. Thus, they are usually unable to guarantee optimality and not even able to provide an assessment of the distance from the optimal solution. In this work, we address the challenge of planning a fault tolerant network through an exact approach, by introducing innovative Mixed-Integer Linear Programming models designed for the planning of meshed distribution networks with loop-feeder or open-loop topology. Differently from other techniques, our approach simplifies the formulation by avoiding the need for fault scenarios, significantly reducing the computational burden of the optimization problem. The outcomes of our approach are the generation of optimal meshed network, which effectively balance cost and reliability of the electric distribution system. Comprehensive studies on realistic test instances show the advantages of the proposed formulations.}, archive = {J_COR}, author = {Renato Bruni and Alberto Geri and Marco Maccioni and Ludovico Nati}, doi = {10.1016/j.cor.2025.107248}, journal = {Computers & Operations Research}, month = {1}, pages = {107248}, shortjournal = {Comput. Oper. Res.}, title = {Optimal planning of power distribution networks with fault-tolerant configuration}, volume = {185}, year = {2026}, }
• (2026). Centrality measures and opinion dynamics in two-layer networks with replica nodes. COR, 185, 107245. (www)

  @article{ , abstract = {We propose two fast and accurate algorithms to approximate game-theoretic centrality measures and examine connection between centrality measures, network properties, and key performance indicators (consensus time and winning rate) of opinion dynamic processes on such networks. As an example, we consider a Zachary’s karate club as a social network and extend it by adding the second (internal) layer of communication. The internal layer represents the network where individuals can share their real opinions with the close friends. The structures of the external and internal layers may be different. The significant positive correlation between internal graph density and consensus time, and significant negative correlation between centrality of authoritative nodes and consensus time are found. The proposed algorithms are verified by a series of experiments from two aspects: the accuracy and the efficiency. The algorithms are novel and can be considered as a contribution to the network theory independently of opinion dynamics as they can be used to calculate node centrality in any weighted graph.}, archive = {J_COR}, author = {Chi Zhao and Elena Parilina}, doi = {10.1016/j.cor.2025.107245}, journal = {Computers & Operations Research}, month = {1}, pages = {107245}, shortjournal = {Comput. Oper. Res.}, title = {Centrality measures and opinion dynamics in two-layer networks with replica nodes}, volume = {185}, year = {2026}, }
• (2026). AI automatic decision in newsvendor model with nash bargaining fairness concern. COR, 185, 107227. (www)

  @article{ , abstract = {This paper investigates the impact of artificial intelligence (AI) automatic ordering and producing decisions on fairness-concerned supply chains under the newsvendor model. We develop a dyadic supply chain model in which the manufacturer acts as the Stackelberg leader while the retailer serves as the follower in a push supply chain. In contrast, their roles are switched in a pull supply chain. We assume that only human decision-making leads to decision regret behavior, whereas AI-automated decision-making does not. Without adopting AI, our results show that fairness concern does not necessarily lead to a decreasing quantity in ordering or producing, which is different from most previous studies. Different from the prior findings, our work reveals that in binding equilibrium, if fairness concerns are considered, the order quantity will decrease, while in non-binding equilibrium, the order quantity may not necessarily be less than the previous results. Interestingly, when decision regret bias is considered for fairness-concerned decision-makers, we can obtain quantity coordination solutions for supply chains under specific conditions. With adopting AI, our results show that increasing fairness concerns are beneficial for improving the follower’s profit while at the expense of sacrificing the leader’s profit margins, while the leader can only benefit from AI adoption when the decision regret bias of the follower is relatively high. It is noteworthy that under certain conditions, AI automation may negatively impact the profits of both push and pull decentralized supply chains. For instance, in low-margin profit scenarios where decision-makers exhibit moderate regret bias and fairness concerns, such effects can emerge. This indicates that under specific circumstances, the human behavioral factors — regret bias and fairness concerns — may sometimes enhance the performance of decentralized supply chain members. Our research findings provide significant practical implications for the adoption of AI-automated decision-making in real-world supply chains.}, archive = {J_COR}, author = {Rui Hou and Yishen Cen and Jianxin Chen}, doi = {10.1016/j.cor.2025.107227}, journal = {Computers & Operations Research}, month = {1}, pages = {107227}, shortjournal = {Comput. Oper. Res.}, title = {AI automatic decision in newsvendor model with nash bargaining fairness concern}, volume = {185}, year = {2026}, }
• (2026). Green horizons: Sustainable global logistics in dynamic supply chain management. COR, 185, 107226. (www)

  @article{ , abstract = {Supply chain management in a global scale involves addressing numerous uncertainties, from demand fluctuations to unforeseen disruptions. Developing advanced solution approaches is critical to manage such complexities and ensure resilience. This study presents a multi-stage stochastic–dynamic model for the global supply chain, incorporating hedging policies. The aim is to identify optimal order scheduling for bill of materials, production planning, and inventory management across warehouses (i.e., materials and finished products). Due to the dynamic nature of the global supply chain (e.g., demand fluctuations, disruptions, and lead time), a multi-stage stochastic model is developed for the stochastic–dynamic supply chain network. To address dynamic factors of real-world global supply chain, an accelerated parallel stochastic dual dynamic integer programming (SDDiP) approach is proposed to deal with disruptions (e.g., political unrest, natural disasters, and pandemics), enhancing supply chain resiliency. To validate the proposed parallel SDDiP , various scenarios with different sizes are generated using the case study and compared to the SDDiP with Benders cuts and integrated stage-wise Lagrangian dual cut ( SWLDC ) (i.e., SDDiP-SWLDC ). According to the obtained results, the proposed parallel node strategy for accelerated SDDiP consistently outperforms the basic stochastic dual dynamic programming (SDDP) and demonstrated robust CPU scalability. Evaluation across various scenario sizes shows stochastic dual dynamic integer programming-mixed integer rounding cuts ( SDDiP-MIR ) achieving faster computation and a smaller 7% optimality gap compared to SDDiP-SWLDC and SDDiP in large-size instances, highlighting its superior performance in complex supply chain settings.}, archive = {J_COR}, author = {Mahsa Mohammadi and Babak Mohamadpour Tosarkani}, doi = {10.1016/j.cor.2025.107226}, journal = {Computers & Operations Research}, month = {1}, pages = {107226}, shortjournal = {Comput. Oper. Res.}, title = {Green horizons: Sustainable global logistics in dynamic supply chain management}, volume = {185}, year = {2026}, }

CSDA - 5

• (2026). Resampling NANCOVA: Nonparametric analysis of covariance in small samples. CSDA, 215, 108290. (www)

  @article{ , abstract = {Analysis of covariance is a crucial method for improving precision of statistical tests for factor effects in randomized experiments. However, existing solutions suffer from one or more of the following limitations: (i) they are not suitable for ordinal data (as endpoints or explanatory variables); (ii) they require semiparametric model assumptions; (iii) they are inapplicable to small data scenarios due to often poor type-I error control; or (iv) they provide only approximate testing procedures and (asymptotically) exact test are missing. A resampling approach to the NANCOVA framework is investigated. NANCOVA is a fully nonparametric model based on relative effects that allows for an arbitrary number of covariates and groups, where both outcome variable (endpoint) and covariates can be metric or ordinal. Novel NANCOVA tests and a nonparametric competitor test without covariate adjustment were evaluated in extensive simulations. Unlike approximate tests in the NANCOVA framework, the proposed resampling version showed good performance in small sample scenarios and maintained the nominal type-I error well. Resampling NANCOVA also provided consistently high power: up to 26 % higher than the test without covariate adjustment in a small sample scenario with 4 groups and two covariates. Moreover, it is shown that resampling NANCOVA provides an asymptotically exact testing procedure, which makes it the first one with good finite sample performance in the present NANCOVA framework. In summary, resampling NANCOVA can be considered a viable tool for analysis of covariance overcoming issues (i) - (iv).}, archive = {J_CSDA}, author = {Konstantin Emil Thiel and Paavo Sattler and Arne C. Bathke and Georg Zimmermann}, doi = {10.1016/j.csda.2025.108290}, journal = {Computational Statistics & Data Analysis}, month = {3}, pages = {108290}, shortjournal = {Comput. Stat. Data Anal.}, title = {Resampling NANCOVA: Nonparametric analysis of covariance in small samples}, volume = {215}, year = {2026}, }
• (2026). Bootstrap-based goodness-of-fit test for parametric families of conditional distributions. CSDA, 215, 108289. (www)

  @article{ , abstract = {A consistent goodness-of-fit test for distributional regression is introduced. The test statistic is based on a process that traces the difference between a nonparametric and a semi-parametric estimate of the marginal distribution function of Y . As its asymptotic null distribution is not distribution-free, a parametric bootstrap method is used to determine critical values. Empirical results suggest that, in certain scenarios, the test outperforms existing specification tests by achieving a higher power and thereby offering greater sensitivity to deviations from the assumed parametric distribution family. Notably, the proposed test does not involve any hyperparameters and can easily be applied to individual datasets using the gofreg-package in R.}, archive = {J_CSDA}, author = {Gitte Kremling and Gerhard Dikta}, doi = {10.1016/j.csda.2025.108289}, journal = {Computational Statistics & Data Analysis}, month = {3}, pages = {108289}, shortjournal = {Comput. Stat. Data Anal.}, title = {Bootstrap-based goodness-of-fit test for parametric families of conditional distributions}, volume = {215}, year = {2026}, }
• (2026). Fast autoregressive model for multivariate dependent outcomes with application to lipidomics analysis for alzheimer’s disease and APOE-ε4. CSDA, 215, 108280. (www)

  @article{ , abstract = {Association analysis of multivariate omics outcomes is challenging due to the high dimensionality and inter-correlation among outcome variables. In practice, the classic multi-univariate analysis approaches are commonly employed, utilizing linear regression models for each individual outcome followed by adjustments for multiplicity through control of the false discovery rate (FDR) or family-wise error rate (FWER). While straightforward, these multi-univariate methods overlook dependencies between outcome variables. This oversight leads to less accurate statistical inferences, characterized by lower power and an increased false discovery rate, ultimately resulting in reduced replicability across studies. Recently, advanced frequentist and Bayesian methods have been developed to account for these dependencies. However, these methods often pose significant computational challenges for researchers in the field. To bridge this gap, a computationally efficient autoregressive multivariate regression model is proposed that explicitly accounts for the dependence structure among outcome variables. Through extensive simulations, it is demonstrated that the approach provides more accurate multivariate inferences than traditional methods and remains robust even under model misspecification. Additionally, the proposed method is applied to investigate whether the associations between serum lipidomics outcomes and Alzheimer’s disease differentiate in ε 4 allele carriers and non-carriers of the apolipoprotein E (APOE) gene.}, archive = {J_CSDA}, author = {Hwiyoung Lee and Zhenyao Ye and Chixiang Chen and Peter Kochunov and L. Elliot Hong and Shuo Chen}, doi = {10.1016/j.csda.2025.108280}, journal = {Computational Statistics & Data Analysis}, month = {3}, pages = {108280}, shortjournal = {Comput. Stat. Data Anal.}, title = {Fast autoregressive model for multivariate dependent outcomes with application to lipidomics analysis for alzheimer’s disease and APOE-ε4}, volume = {215}, year = {2026}, }
• (2026). Change-point detection in regression models via the max-EM algorithm. CSDA, 215, 108278. (www)

  @article{ , abstract = {The problem of breakpoint detection is considered within a regression modeling framework. A novel method, the max-EM algorithm, is introduced, combining a constrained Hidden Markov Model with the Classification-EM algorithm. This algorithm has linear complexity and provides accurate detection of breakpoints and estimation of parameters. A theoretical result is derived, showing that the likelihood of the data, as a function of the regression parameters and the breakpoints location, increases at each step of the algorithm. Two initialization methods for the breakpoints location are also presented to address local maxima issues. Finally, a statistical test in the one breakpoint situation is developed. Simulation experiments based on linear, logistic, Poisson and Accelerated Failure Time regression models show that the final method that includes the initialization procedure and the max-EM algorithm has a strong performance both in terms of parameters estimation and breakpoints detection. The statistical test is also evaluated and exhibits a correct rejection rate under the null hypothesis and a strong power under various alternatives. Two real dataset are analyzed, the UCI bike sharing and the health disease data, where the interest of the method to detect heterogeneity in the distribution of the data is illustrated.}, archive = {J_CSDA}, author = {Modibo Diabaté and Grégory Nuel and Olivier Bouaziz}, doi = {10.1016/j.csda.2025.108278}, journal = {Computational Statistics & Data Analysis}, month = {3}, pages = {108278}, shortjournal = {Comput. Stat. Data Anal.}, title = {Change-point detection in regression models via the max-EM algorithm}, volume = {215}, year = {2026}, }
• (2026). Recursive nonparametric predictive for a discrete regression model. CSDA, 215, 108275. (www)

  @article{ , abstract = {A recursive algorithm is proposed to estimate a set of distribution functions indexed by a regressor variable. The procedure is fully nonparametric and has a Bayesian motivation and interpretation. Indeed, the recursive algorithm follows a certain Bayesian update, defined by the predictive distribution of a Dirichlet process mixture of linear regression models. Consistency of the algorithm is demonstrated under mild assumptions, and numerical accuracy in finite samples is shown via simulations and real data examples. The algorithm is very fast to implement, it is parallelizable, sequential, and requires limited computing power.}, archive = {J_CSDA}, author = {Lorenzo Cappello and Stephen G. Walker}, doi = {10.1016/j.csda.2025.108275}, journal = {Computational Statistics & Data Analysis}, month = {3}, pages = {108275}, shortjournal = {Comput. Stat. Data Anal.}, title = {Recursive nonparametric predictive for a discrete regression model}, volume = {215}, year = {2026}, }

CVIU - 18

• (2025). IP-CAM: Class activation mapping based on importance weights and principal-component weights for better and simpler visual explanations. CVIU, 261, 104523. (www)

  @article{ , abstract = {Visual explanations of deep neural networks (DNNs) have gained considerable importance in deep learning due to the lack of interpretability, which constrains human trust in DNNs. This paper proposes a new gradient-free class activation map (CAM) architecture called importance principal-component CAM (IP-CAM). The architecture not only improves the prediction accuracy of networks but also provides simpler and more reliable visual explanations. It adds importance weight layers before the classifier and assigns an importance weight to each activation map. After fine-tuning, it selects images with the highest prediction score for each class, performs principal component analysis (PCA) on activation maps of all channels, and regards the eigenvector of the first principal component as principal-component weights for that class. The final saliency map is obtained by linearly combining the activation maps, importance weights and principal-component weights. IP-CAM is evaluated on the ILSVRC 2012 dataset and RSD46-WHU dataset, whose results show that IP-CAM performs better than most previous CAM variants in recognition and localization tasks. Finally, the method is applied as a tool for interpretability, and the results illustrate that IP-CAM effectively unveils the decision-making process of DNNs through saliency maps.}, archive = {J_CVIU}, author = {Wenyi Zhang and Haoran Zhang and Xisheng Zhang and Xiaohua Shen and Lejun Zou}, doi = {10.1016/j.cviu.2025.104523}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104523}, shortjournal = {Comput. Vis. Image Understanding}, title = {IP-CAM: Class activation mapping based on importance weights and principal-component weights for better and simpler visual explanations}, volume = {261}, year = {2025}, }
• (2025). MSBATN: Multi-stage boundary-aware transformer network for action segmentation in untrimmed surgical videos. CVIU, 261, 104521. (www)

  @article{ , abstract = {Understanding actions within surgical workflows is critical for evaluating post-operative outcomes and enhancing surgical training and efficiency. Capturing and analysing long sequences of actions in surgical settings is challenging due to the inherent variability in individual surgeon approaches, which are shaped by their expertise and preferences. This variability complicates the identification and segmentation of distinct actions with ambiguous boundary start and end points. The traditional models, such as MS-TCN, which rely on large receptive fields, cause over-segmentation or under-segmentation, where distinct actions are incorrectly aligned. To address these challenges, we propose the Multi-Stage Boundary-Aware Transformer Network (MSBATN) with hierarchical sliding window attention to improve action segmentation. Our approach effectively manages the complexity of varying action durations and subtle transitions by accurately identifying start and end action boundaries in untrimmed surgical videos. MSBATN introduces a novel unified loss function that optimises action classification and boundary detection as interconnected tasks. Unlike conventional binary boundary detection methods, our innovative boundary weighing mechanism leverages contextual information to precisely identify action boundaries. Extensive experiments on three challenging surgical datasets demonstrate that MSBATN achieves state-of-the-art performance, with superior F1 scores at 25% and 50% thresholds and competitive results across other metrics.}, archive = {J_CVIU}, author = {Rezowan Shuvo and M.S. Mekala and Eyad Elyan}, doi = {10.1016/j.cviu.2025.104521}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104521}, shortjournal = {Comput. Vis. Image Understanding}, title = {MSBATN: Multi-stage boundary-aware transformer network for action segmentation in untrimmed surgical videos}, volume = {261}, year = {2025}, }
• (2025). SVFFNet: A scale-aware voxel flow fusion network for video prediction. CVIU, 261, 104520. (www)

  @article{ , abstract = {Video prediction is a challenging task due to the potential for various motion scales in the complex scene. The diversity of motion scales stems from the time-variant and object-dependent motion magnitudes, as well as the multiple image resolutions across datasets. However, the vast majority of frame forecasting networks do not distinguish between treatment of different motion scales. Therefore, their receptive field is normally insufficient to capture larger-scale motions. Those that do, often yield significant local distortions in the predicted images. The reasons lie in their fixed choice of scale factors and lack of cross-scale interaction between motion features. In this work, we propose a Scale-Aware Voxel Flow Fusion Network (SVFFNet) to address the motion scale inconsistency problem and fully integrate multi-scale feature. This network consists of a set of flow estimation blocks, each block containing a selector module and a fusion module. The selector module adaptively selects the appropriate scale-processing branch for the input frames, thus facilitating acquisition of more refined features for large-scale motion. The fusion module then combines these features with the original motion information via an attention mechanism, preserving the actually existing structural details. Experimental results on four widely used benchmark datasets demonstrate that our method outperforms previously published baselines for video prediction. The code is available at: https://github.com/zyaojlu/SVFFNet .}, archive = {J_CVIU}, author = {Yao Zhou and Jinpeng Wei and Xueyong Zhang and Yusong Zhai and Jian Wei}, doi = {10.1016/j.cviu.2025.104520}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104520}, shortjournal = {Comput. Vis. Image Understanding}, title = {SVFFNet: A scale-aware voxel flow fusion network for video prediction}, volume = {261}, year = {2025}, }
• (2025). Generalization-preserving adaptation of vision-language models for open-vocabulary segmentation. CVIU, 261, 104518. (www)

  @article{ , abstract = {Recent progress in large-scale Vision-Language Models (VLMs) has significantly advanced open-vocabulary segmentation. Previous works typically either generate class-agnostic masks and classify them with frozen VLMs, or align the mask generator features with VLM text features. These approaches face challenges of weak spatial discrimination ability of frozen VLMs and poor generalization due to unreliable vision-language alignment. This paper introduces a novel Generalization-Preserving Adaptation (GPA) of VLMs for open-vocabulary segmentation. GPA enhances the spatial discrimination capability of pre-trained VLMs through an efficient fine-tuning scheme, which incorporates a spatial adaptation module comprising spatial dependency modeling and low-rank feature modulation for preserving the feature space. Additionally, GPA proposes a context-aware feature aggregation module to extract mask features better aligned with the VLM features for mask classification. It performs decoupled context modeling that generates object-agnostic contextualized feature map and object-specific classification maps for accentuating discriminative and contextual clues. By maintaining the original VLM feature distribution for vision-language alignment, GPA effectively preserves the generalization capabilities of VLMs while enhancing segmentation performance. Extensive experiments on multiple open-vocabulary panoptic and semantic segmentation benchmarks demonstrate both superior effectiveness and generalization capabilities compared to previous works.}, archive = {J_CVIU}, author = {Zhen Chen and Hao Tang and Shiliang Zhang}, doi = {10.1016/j.cviu.2025.104518}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104518}, shortjournal = {Comput. Vis. Image Understanding}, title = {Generalization-preserving adaptation of vision-language models for open-vocabulary segmentation}, volume = {261}, year = {2025}, }
• (2025). Beyond geometry: The power of texture in interpretable 3D person ReID. CVIU, 261, 104517. (www)

  @article{ , abstract = {This paper presents FusionTexReIDNet, a robust framework for 3D person re-identification that uniquely leverages UVTexture to enhance both performance and explainability. Unlike existing 3D person ReID approaches that simply overlay textures on point clouds, our method exploits the full potential of UVTexture through its high resolution and normalized coordinate properties. The framework consists of two main streams: a UVTexture stream that processes appearance features and a 3D stream that handles geometric information. These streams are fused through an effective combination of KNN, attribute-based, and explainable re-ranking strategies. Our approach introduces explainability to 3D person ReID through the visualization of activation maps on UVTextures, providing insights into the model’s decision-making process by highlighting discriminative regions. By incorporating the Intersection-Alignment Score derived from activation maps and visible clothing masks, we further improve the ReID accuracy. Extensive experiments demonstrate that FusionTexReIDNet achieves state-of-the-art performance across various scenarios, with Rank-1 accuracies of 98.5% and 89.7% Rank-1 on benchmark datasets, while providing interpretable results through its explainable component.}, archive = {J_CVIU}, author = {Huy Nguyen and Kien Nguyen and Akila Pemasiri and Sridha Sridharan and Clinton Fookes}, doi = {10.1016/j.cviu.2025.104517}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104517}, shortjournal = {Comput. Vis. Image Understanding}, title = {Beyond geometry: The power of texture in interpretable 3D person ReID}, volume = {261}, year = {2025}, }
• (2025). Synergistic dual and efficient additive attention network for no-reference image quality assessment. CVIU, 261, 104516. (www)

  @article{ , abstract = {No-Reference Image Quality Assessment (NR-IQA) aims to evaluate the perceptual quality of images in alignment with human subjective judgments. However, most existing NR-IQA methods, while striving for high accuracy, often neglect computational complexity. To address this challenge, we propose a Synergistic Spatial and Channel and Efficient Additive Attention Network for NR-IQA. In our approach, we first employ a feature extraction module to derive features rich in both distortion and semantic information. Subsequently, we introduce a spatial-channel synergistic attention mechanism to enhance feature representations across spatial and channel dimensions. This attention module focuses on the most salient regions of the image and modulates feature responses accordingly, enabling the network to emphasize critical distortions and semantic features pertinent to perceptual quality assessment. Specifically, the spatial attention mechanism identifies significant regions that substantially contribute to quality perception, while the channel attention mechanism adjusts the importance of each feature channel, ensuring effective utilization of spatial and channel-specific information. Furthermore, to enhance the model’s robustness, we incorporate an Efficient Additive Attention mechanism alongside a Multi-scale Feed-forward Network, designed to reduce computational costs without compromising performance. Finally, a dual-branch structure for patch-weighted quality prediction is employed to derive the final quality score based on the weighted scores of individual patches. Extensive experimental evaluations on four widely used benchmark datasets demonstrate that the proposed method surpasses several state-of-the-art NR-IQA approaches in both performance and computational efficiency.}, archive = {J_CVIU}, author = {Zhou Fang and Baiming Feng and Ning Li}, doi = {10.1016/j.cviu.2025.104516}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104516}, shortjournal = {Comput. Vis. Image Understanding}, title = {Synergistic dual and efficient additive attention network for no-reference image quality assessment}, volume = {261}, year = {2025}, }
• (2025). Iterative caption generation with heuristic guidance for enhancing knowledge-based visual question answering. CVIU, 261, 104515. (www)

  @article{ , abstract = {The advent of large language models (LLMs) has significantly advanced Knowledge-based Visual Question Answering (KBVQA) by reducing the reliance on external knowledge bases. Traditional methods often generate captions in a single pass, which can struggle with complex questions due to difficulty in precisely identifying key visual components. This challenge undermines the reasoning capabilities of LLMs, which require accurate, semantically aligned captions to answer complex questions effectively. To address this limitation, we propose ICGHG I terative C aption G eneration with H euristic G uidance, a novel framework that refines captions iteratively. Our approach incorporates a dynamic loop where captions are continuously refined based on heuristic feedback from a set of candidate answers and the question itself, ensuring that the final caption provides accurate semantic alignment with both the visual content and the question. By leveraging this iterative process, ICGHG mitigates common issues such as hallucinations and improves the quality of the generated captions. Extensive experiments on OK-VQA, A-OKVQA, and FVQA datasets demonstrate that ICGHG significantly outperforms existing methods, achieving 57.5%, 60.2%, and 69.4% accuracy on their respective test sets, setting new benchmarks in KBVQA accuracy.}, archive = {J_CVIU}, author = {Fengyuan Liu and Zhongjian Hu and Peng Yang and Xingyu Liu}, doi = {10.1016/j.cviu.2025.104515}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104515}, shortjournal = {Comput. Vis. Image Understanding}, title = {Iterative caption generation with heuristic guidance for enhancing knowledge-based visual question answering}, volume = {261}, year = {2025}, }
• (2025). KD-mamba: Selective state space models with knowledge distillation for trajectory prediction. CVIU, 261, 104499. (www)

  @article{ , abstract = {Trajectory prediction is a key component of intelligent mobility systems and human–robot interaction. The inherently stochastic nature of human behavior, coupled with external environmental influences, poses significant challenges for long-term prediction. However, existing approaches struggle to effectively model spatial interactions and accurately predict long-term destinations, while their high computational demands limit real-world applicability. To address these limitations, this paper presents KD-Mamba, the Selective State Space Models with Knowledge Distillation for trajectory prediction. The model incorporates the U-CMamba module, which features a U-shaped encoder–decoder architecture. By integrating convolutional neural networks (CNN) with the Mamba mechanism, this module effectively captures local spatial interactions and global contextual information of human motion patterns. Subsequently, we introduce a Bi-Mamba module, which captures long-term dependencies in human movement, ensuring a more accurate representation of trajectory dynamics. Knowledge distillation strengthens both modules by facilitating knowledge transfer across diverse scenarios. Compared to transformer-based approaches, KD-Mamba reduces computational complexity from quadratic to linear. Extensive experimental results from two real-world trajectory datasets indicate that KD-Mamba outperforms the existing mainstream baselines. The proposed method provides insights into the application of trajectory prediction in human-in-the-loop assistive systems.}, archive = {J_CVIU}, author = {Shaokang Cheng and Sourav Das and Shiru Qu and Lamberto Ballan}, doi = {10.1016/j.cviu.2025.104499}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104499}, shortjournal = {Comput. Vis. Image Understanding}, title = {KD-mamba: Selective state space models with knowledge distillation for trajectory prediction}, volume = {261}, year = {2025}, }
• (2025). Gloss-free sign language translation: An unbiased evaluation of progress in the field. CVIU, 261, 104498. (www)

  @article{ , abstract = {Sign Language Translation (SLT) aims to automatically convert visual sign language videos into spoken language text and vice versa. While recent years have seen rapid progress, the true sources of performance improvements often remain unclear. Do reported performance gains come from methodological novelty, or from the choice of a different backbone, training optimizations, hyperparameter tuning, or even differences in the calculation of evaluation metrics? This paper presents a comprehensive study of recent gloss-free SLT models by re-implementing key contributions in a unified codebase. We ensure fair comparison by standardizing preprocessing, video encoders, and training setups across all methods. Our analysis shows that many of the performance gains reported in the literature often diminish when models are evaluated under consistent conditions, suggesting that implementation details and evaluation setups play a significant role in determining results. We make the codebase publicly available here 1 to support transparency and reproducibility in SLT research.}, archive = {J_CVIU}, author = {Ozge Mercanoglu Sincan and Jian He Low and Sobhan Asasi and Richard Bowden}, doi = {10.1016/j.cviu.2025.104498}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104498}, shortjournal = {Comput. Vis. Image Understanding}, title = {Gloss-free sign language translation: An unbiased evaluation of progress in the field}, volume = {261}, year = {2025}, }
• (2025). Two-stage attribute-guided dual attention network for fine-grained fashion retrieval. CVIU, 261, 104497. (www)

  @article{ , abstract = {Fine-grained clothing retrieval is essential for intelligent shopping and personalized recommendation systems. However, conventional methods often fail to capture subtle attribute variations. This paper proposes a novel two-stage attribute-guided dual attention network. The network combines global and local feature extraction with Attribute-aware Multi-Scale Spatial Attention (AMSA) and Attribute-guided Dynamic Channel Attention (ADCA). AMSA captures attribute-specific spatial details at multiple scales, while ADCA dynamically adjusts channel importance based on attribute embeddings, enabling precise attribute-level similarity modeling. A multi-level joint loss function further optimizes both global and local representations and enhances feature alignment. Experiments on FashionAI and the self-built FGDress dataset show that the proposed method achieves mAP scores of 66.01% and 73.98%, respectively, outperforming baseline approaches. Attribute-level analysis confirms robust recognition of both well-defined and challenging attributes. These results validate the practicality and generalizability of the proposed framework, with promising applications in personalized recommendation, fashion trend analysis, and design evaluation.}, archive = {J_CVIU}, author = {Bo Pan and Jun Xiang and Ning Zhang and Ruru Pan}, doi = {10.1016/j.cviu.2025.104497}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104497}, shortjournal = {Comput. Vis. Image Understanding}, title = {Two-stage attribute-guided dual attention network for fine-grained fashion retrieval}, volume = {261}, year = {2025}, }
• (2025). JSF: A joint spatial-frequency domain network for low-light image enhancement. CVIU, 261, 104496. (www)

  @article{ , abstract = {The enhancement of low-light images remains a prominent focus in the field of image processing. The degree of lightness significantly influences vision-based intelligent recognition and analysis. Departing from conventional methods, this paper proposes an innovative joint spatial-frequency domain network for low-light image enhancement, referred to as JSF. In the spatial domain, brightness is optimized through the amalgamation of global and local information. In the frequency domain, noise is reduced and details are amplified using Fourier Transformation to carry out amplitude and phase enhancement. Additionally, the enhanced results from the aforementioned domains are fused by linear and nonlinear stretching. To validate the effectiveness of JSF, this paper presents both qualitative and quantitative comparison results, demonstrating its superiority over several existing state-of-the-art methods.}, archive = {J_CVIU}, author = {Yahong Wu and Feng Liu and Rong Wang}, doi = {10.1016/j.cviu.2025.104496}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104496}, shortjournal = {Comput. Vis. Image Understanding}, title = {JSF: A joint spatial-frequency domain network for low-light image enhancement}, volume = {261}, year = {2025}, }
• (2025). FCNet: A feature complementary network for nighttime flare removal. CVIU, 261, 104495. (www)

  @article{ , abstract = {Nighttime image flare removal is a very challenging task due to the presence of various types of unfavorable degrading effects, including glare, shimmer, streak and saturated blobs. Most of the existing methods focus on the spatial domain and limited perception field, resulting in incomplete flare removal and severe artifacts. To address these challenges, we propose a two-stage feature complementary network for nighttime flare removal, which is used for flare perception and removal, respectively. In the first stage, a Spatial-Frequency Complementary Module (SFCM) is designed to perceive the flare region from different domains to get a mask of the flare. In the second stage, the flare mask and image are fed into the Spatial-Frequency Complementary Gating Module (SFCGM) to preserve the background information, while removing the flares from different angles and restoring the detailed features. Finally the flare and non-flare regions are modeled by the Flare Interactive Module (FIM) to refine the flare regions at a fine-grained level to suppress the artifact problem. Extensive experiments on Flare 7K++ validate the superiority of the proposed approach over state-of-the-arts, both qualitatively and quantitatively.}, archive = {J_CVIU}, author = {Kejing Qi and Bo Wang and Chongyi Li}, doi = {10.1016/j.cviu.2025.104495}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104495}, shortjournal = {Comput. Vis. Image Understanding}, title = {FCNet: A feature complementary network for nighttime flare removal}, volume = {261}, year = {2025}, }
• (2025). GRE-net: A forgery image detection framework based on gradient feature and reconstruction error. CVIU, 261, 104494. (www)

  @article{ , abstract = {With the continuous technological breakthroughs in Generative Adversarial Networks (GANs) and diffusion models, remarkable progress has been achieved in the field of image generation. These technologies enable the creation of highly realistic images, thereby intensifying the risk of spreading fake information. However, traditional image detectors face a growing challenge of inadequate generalization capabilities when confronted with images generated by models that were not included during the training phase. To tackle this challenge, we introduce a novel detection framework, named GRE-Net (Network integrating Gradient and Reconstruction Error), which extracts gradient feature through the DPG module and calculates the reconstruction error utilizing the DIRE method. By integrating these two aspects into a comprehensive feature representation, GRE-Net effectively detects the authenticity of images. Specifically, we devise a dual-branch model that leverages the proposed DPG (Discriminator of ProjectedGAN to extract Gradient) module to extract gradient feature from images and concurrently employs the DIRE (DIffusion Reconstruction Error) method to obtain the diffusion reconstruction error of images. By fusing the features extracted from these two modules as a universal representation, we describe the artifacts produced by generative models, crafting a comprehensive detector capable of identifying both GAN-generated and diffusion model-generated images. Notably, the DPG approach utilizes the discriminator of ProjectedGAN as an intermediary bridge, mapping all data into the gradient domain. This transformation process effectively captures the intrinsic feature differences during the image generation process. Subsequently, the gradient feature are fed into a classifier to achieve efficient discrimination between authentic and fake images. To validate the efficacy of our proposed detector, we conducted evaluations on a dataset comprising images generated by ten diverse diffusion models and GANs. Extensive experiments demonstrate that our detector exhibits stronger generalization capabilities and higher robustness, rendering it suitable for real-world generated image detection tasks.}, archive = {J_CVIU}, author = {Wenqing Wu and Xinyi Shi and Jinghai Ai and Xiaodong Wang}, doi = {10.1016/j.cviu.2025.104494}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104494}, shortjournal = {Comput. Vis. Image Understanding}, title = {GRE-net: A forgery image detection framework based on gradient feature and reconstruction error}, volume = {261}, year = {2025}, }
• (2025). Mitigating forgetting in the adaptation of CLIP for few-shot classification. CVIU, 261, 104493. (www)

  @article{ , abstract = {Adapter-style efficient transfer learning has demonstrated outstanding performance in fine-tuning vision-language models, especially in scenarios with limited data. However, existing methods fail to effectively balance the prior knowledge acquired during the pre-training process and the training samples. To address this problem, we propose a method called Mitigating Forgetting in the Adaptation (MiFA) of CLIP. MiFA first employs class prototypes to represent the most prominent features of a class, and these prototypes provide a robust initialization for the classifier. To overcome the forgetting of prior knowledge, MiFA then leverages a memory module that retains the initial parameters and the parameters of training history by creating a memory weight through momentum. The weight is used to initialize a new classification layer, which, along with the original layer, guides each other to balance prior knowledge and feature adaptation. Similarly, in the text processing branch, a parallel initialization strategy is adopted to ensure that the model’s performance is improved. Text features are employed to initialize a text classification layer, and CLIP logits help prevent excessive forgetting of useful text information. Extensive experiments have demonstrated the effectiveness of our method.}, archive = {J_CVIU}, author = {Jiale Cao and Yuanheng Liu and Zhong Ji and Jingren Liu and Aiping Yang and Yanwei Pang}, doi = {10.1016/j.cviu.2025.104493}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104493}, shortjournal = {Comput. Vis. Image Understanding}, title = {Mitigating forgetting in the adaptation of CLIP for few-shot classification}, volume = {261}, year = {2025}, }
• (2025). Joint multi-dimensional dynamic attention and transformer for general image restoration. CVIU, 261, 104491. (www)

  @article{ , abstract = {Outdoor images often suffer from severe degradation due to rain, haze, and noise, impairing image quality and challenging high-level tasks. Current image restoration methods struggle to handle complex degradation while maintaining efficiency. This paper introduces a novel image restoration architecture that combines multi-dimensional dynamic attention and self-attention within a U-Net framework. To leverage the global modeling capabilities of transformers and the local modeling capabilities of convolutions, we integrate sole CNNs in the encoder–decoder and sole transformers in the latent layer. Additionally, we design convolutional kernels with selected multi-dimensional dynamic attention to capture diverse degraded inputs efficiently. A transformer block with transposed self-attention further enhances global feature extraction while maintaining efficiency. Extensive experiments demonstrate that our method achieves a better balance between performance and computational complexity across five image restoration tasks: deraining, deblurring, denoising, dehazing, and enhancement, as well as superior performance for high-level vision tasks. The source code will be available at https://github.com/House-yuyu/MDDA-former .}, archive = {J_CVIU}, author = {Huan Zhang and Xu Zhang and Nian Cai and Jianglei Di and Yun Zhang}, doi = {10.1016/j.cviu.2025.104491}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104491}, shortjournal = {Comput. Vis. Image Understanding}, title = {Joint multi-dimensional dynamic attention and transformer for general image restoration}, volume = {261}, year = {2025}, }
• (2025). LAM-YOLO: Drones-based small object detection on lighting-occlusion attention mechanism YOLO. CVIU, 261, 104489. (www)

  @article{ , abstract = {Drone-based target detection presents inherent challenges, including the high density and overlap of targets in drone images, as well as the blurriness of targets under varying lighting conditions, which complicates accurate identification. Traditional methods often struggle to detect numerous small, densely packed targets against complex backgrounds. To address these challenges, we propose LAM-YOLO, an object detection model specifically designed for drone-based applications. First, we introduce a light-occlusion attention mechanism to enhance the visibility of small targets under diverse lighting conditions. Additionally, we incorporate Involution modules to improve feature layer interactions. Second, we employ an improved SIB-IoU as the regression loss function to accelerate model convergence and enhance localization accuracy. Finally, we implement a novel detection strategy by introducing two auxiliary detection heads to better identify smaller-scale targets. Our quantitative results demonstrate that LAM-YOLO outperforms methods such as Faster R-CNN, YOLOv11, and YOLOv12 in terms of mAP@0.5 and mAP@0.5:0.95 on the VisDrone2019 public dataset. Compared to the original YOLOv8, the average precision increases by 7.1%. Additionally, the proposed SIB-IoU loss function not only accelerates convergence speed during training but also improves average precision compared to the traditional loss function.}, archive = {J_CVIU}, author = {Yuchen Zheng and Yuxin Jing and Jufeng Zhao and Guangmang Cui}, doi = {10.1016/j.cviu.2025.104489}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104489}, shortjournal = {Comput. Vis. Image Understanding}, title = {LAM-YOLO: Drones-based small object detection on lighting-occlusion attention mechanism YOLO}, volume = {261}, year = {2025}, }
• (2025). Camera pose in SfT and NRSfM under isometric and weaker deformation models. CVIU, 261, 104488. (www)

  @article{ , abstract = {Camera pose is a very natural concept in 3D vision in the rigid setting. It is however much more difficult to work with in deformable settings. Consequently, numerous deformable reconstruction methods simply ignore camera pose. We analyse the concept of pose in deformable settings and prove that it is unconstrained with the existing formulations, properly justifying the existing pose-less methods reconstructing structure only. We explain this result intuitively by the impossibility to define an intrinsic coordinate frame to a general deforming object. The proposed analysis uses the isometric deformation model and extends to the weaker models including conformality and equiareality We propose a novel prior to rescue camera pose estimation in deformable settings, which attributes the deforming object’s dominant rigid-body motion to the camera. We show that adding this prior to any existing formulation fully constrains camera pose and leads to elegant two-step solution methods, involving deformable structure reconstruction using a base method in the first step, and absolute orientation or Procrustes analysis in the second step. We derive the proposed approach for the template-based and template-less settings, respectively implemented using Shape-from-Template (SfT) and Non-Rigid Structure-from-Motion (NRSfM) as base methods and validate them experimentally, showing that the computed pose is qualitatively and quantitatively plausible.}, archive = {J_CVIU}, author = {Adrien Bartoli and Agniva Sengupta}, doi = {10.1016/j.cviu.2025.104488}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104488}, shortjournal = {Comput. Vis. Image Understanding}, title = {Camera pose in SfT and NRSfM under isometric and weaker deformation models}, volume = {261}, year = {2025}, }
• (2025). Bi-granularity balance learning for long-tailed image classification. CVIU, 261, 104469. (www)

  @article{ , abstract = {In long-tailed datasets, the training of deep neural network-based models faces challenges, where the model may become biased towards the head classes with abundant training data, resulting in poor performance on tail classes with limited samples. Most current methods employ contrastive learning to learn more balanced representations by finding the class center. However, these methods use class centers to address local imbalance within a mini-batch, they overlook the global imbalance between batches throughout an epoch, caused by the long-tailed distribution of the dataset. In this paper, we propose bi-granularity balance learning to address the two-layer imbalance. We decouple the attraction–repulsion term in contrastive loss into two independent components: global and local balance. The global balance component focuses on capturing semantic information from different perspectives of the image and shifting learning attention from the head classes to the tail classes in the global perspective. The local balance component aims to learn inter-class separability from the local perspective. The proposed method efficiently learns the intra-class compactness and inter-class separability in long-tailed model training and improves the performance of the long-tailed model. Experimental results show that the proposed method achieves competitive performance on long-tailed benchmarks such as CIFAR-10/100-LT, TinyImageNet-LT, and iNaturalist 2018.}, archive = {J_CVIU}, author = {Ning Ren and Xiaosong Li and Yanxia Wu and Yan Fu}, doi = {10.1016/j.cviu.2025.104469}, journal = {Computer Vision and Image Understanding}, month = {11}, pages = {104469}, shortjournal = {Comput. Vis. Image Understanding}, title = {Bi-granularity balance learning for long-tailed image classification}, volume = {261}, year = {2025}, }

DISOPT - 4

• (2025). Minimizing maximum dissatisfaction in the allocation of indivisible items under a common preference graph. DISOPT, 58, 100913. (www)

  @article{ , abstract = {We consider the task of allocating indivisible items to agents, when the agents’ preferences over the items are identical. The preferences are captured by means of a directed acyclic graph, with vertices representing items and an edge ( a , b ) , meaning that each of the agents prefers item a over item b . The dissatisfaction of an agent is measured by the number of items that the agent does not receive and for which it also does not receive any more preferred item. The aim is to allocate the items to the agents in a fair way, i.e., to minimize the maximum dissatisfaction among the agents. We study the status of computational complexity of that problem and establish the following dichotomy: the problem is NP -hard for the case of at least three agents, even on fairly restricted graphs, but polynomially solvable for two agents. We also provide several polynomial-time results with respect to different underlying graph structures, such as graphs of width at most two and tree-like structures such as stars and matchings. These findings are complemented with fixed parameter tractability results related to path modules and independent set modules. Techniques employed in the paper include bottleneck assignment problem, greedy algorithm, dynamic programming, maximum network flow, and integer linear programming.}, archive = {J_DISOPT}, author = {Nina Chiarelli and Clément Dallard and Andreas Darmann and Stefan Lendl and Martin Milanič and Peter Muršič and Ulrich Pferschy}, doi = {10.1016/j.disopt.2025.100913}, journal = {Discrete Optimization}, month = {11}, pages = {100913}, shortjournal = {Discret. Optim.}, title = {Minimizing maximum dissatisfaction in the allocation of indivisible items under a common preference graph}, volume = {58}, year = {2025}, }
• (2025). Computational aspects of lifted cover inequalities for knapsacks with few different weights. DISOPT, 58, 100912. (www)

  @article{ , abstract = {Cutting planes are frequently used for solving integer programs. A common strategy is to derive cutting planes from building blocks or a substructure of the integer program. In this paper, we focus on knapsack constraints that arise from single row relaxations. Among the most popular classes derived from knapsack constraints are lifted minimal cover inequalities. The separation problem for these inequalities is NP-hard though, and one usually separates them heuristically, therefore not fully exploiting their potential. For many benchmarking instances however, it turns out that many knapsack constraints only have few different coefficients. This motivates the concept of sparse knapsacks where the number of different coefficients is a small constant, independent of the number of variables present. For such knapsacks, we observe that there are only polynomially many different classes of structurally equivalent minimal covers. This opens the door to specialized techniques for using lifted minimal cover inequalities. In this article we will discuss two such techniques, which are based on specialized sorting methods. On the one hand, we present new separation routines that separate equivalence classes of inequalities rather than individual inequalities. On the other hand, we derive compact extended formulations that express all lifted minimal cover inequalities by means of a polynomial number of constraints. These extended formulations are based on tailored sorting networks that express our separation algorithm by linear inequalities. We conclude the article by a numerical investigation of the different techniques for popular benchmarking instances.}, archive = {J_DISOPT}, author = {Christopher Hojny and Cédric Roy}, doi = {10.1016/j.disopt.2025.100912}, journal = {Discrete Optimization}, month = {11}, pages = {100912}, shortjournal = {Discret. Optim.}, title = {Computational aspects of lifted cover inequalities for knapsacks with few different weights}, volume = {58}, year = {2025}, }
• (2025). An improved bound for the price of anarchy for related machine scheduling. DISOPT, 58, 100911. (www)

  @article{ , abstract = {In this paper, we introduce an improved upper bound for the efficiency of Nash equilibria in utilitarian scheduling games on related machines. The machines have varying speeds and adhere to the shortest processing time first policy. The goal of each job is to minimize its completion time, while the social objective is to minimize the sum of completion times. Our main finding establishes an upper bound of 2 − 1 / ( 4 m − 2 ) on the price of anarchy for the general case of m machines. We improve this bound to 3/2 for the case of two machines, and to 2 − 1 / ( 2 m ) for the general case of m machines when the machines have divisible speeds, i.e., if the speed of each machine is divisible by the speed of any slower machine.}, archive = {J_DISOPT}, author = {André Berger and Arman Rouhani and Marc Schröder}, doi = {10.1016/j.disopt.2025.100911}, journal = {Discrete Optimization}, month = {11}, pages = {100911}, shortjournal = {Discret. Optim.}, title = {An improved bound for the price of anarchy for related machine scheduling}, volume = {58}, year = {2025}, }
• (2025). On the circuit diameter conjecture for counterexamples to the hirsch conjecture. DISOPT, 58, 100910. (www)

  @article{ , abstract = {Circuit diameters of polyhedra are a fundamental tool for studying the complexity of circuit augmentation schemes for linear programming and for finding lower bounds on combinatorial diameters. The main open problem in this area is the circuit diameter conjecture, the analogue of the Hirsch conjecture in the circuit setting. A natural question is whether the well-known counterexamples to the Hirsch conjecture carry over. Previously, Stephen and Yusun showed that the Klee-Walkup counterexample to the unbounded Hirsch conjecture does not transfer to the circuit setting. Our main contribution is to show that the original counterexamples for other variants, using monotone walks or for bounded polytopes, also do not transfer. A challenge lies in the dependence of circuit diameters on the specific realization of a polyhedron. We discuss for which realizations, in addition to the original ones from the literature, our tools resolve this question. Our results rely on new observations on structural properties of these counterexamples. To analyze the bounded case, we exploit the geometry of certain 2-faces of the polytopes underlying all known bounded Hirsch counterexamples in Santos’ work. For Todd’s monotone Hirsch counterexample, we study linear programs on spindles and prove sufficient conditions for short monotone circuit walks to exist. We then enumerate all linear programs over Todd’s polytope and find four new orientations that contradict the monotone Hirsch conjecture, while the remaining 7107 satisfy the bound. The conclusion then follows by applying these sufficient conditions to Todd’s counterexample.}, archive = {J_DISOPT}, author = {Alexander E. Black and Steffen Borgwardt and Matthias Brugger}, doi = {10.1016/j.disopt.2025.100910}, journal = {Discrete Optimization}, month = {11}, pages = {100910}, shortjournal = {Discret. Optim.}, title = {On the circuit diameter conjecture for counterexamples to the hirsch conjecture}, volume = {58}, year = {2025}, }

DKE - 16

• (2026). Temporal knowledge graph recommendation with sequence-aware and path reasoning. DKE, 161, 102522. (www)

  @article{ , abstract = {Knowledge graph recommendation (KGRec) models not only alleviate the issues of data sparsity and the cold start problem encountered by traditional models but also enhance interpretability and credibility through the provision of explicit recommendation rationales. Nonetheless, existing KGRec models predominantly concentrate on extracting static structural features of user preferences from KG, often neglecting the dynamic temporal features, such as purchase time and click time. This oversight results in considerable limitations in recommendation performance. In response to this challenge, this paper introduces a novel temporal knowledge graph recommendation model (TKGRec), which fully utilizes both dynamic temporal feature and static structure feature for better recommendation. We specifically construct a temporal KG that encapsulates both static and dynamic user–item interactions. Based on the new environment, we propose a sequence-aware and path reasoning framework, in which the sequence-aware module employs a dual-attention mechanism to distill temporal features from interactions, whereas the path reasoning module utilizes reinforcement learning to extract path features. These two modules are seamlessly fused and iteratively refined to capture a more holistic understanding of user preferences. Experimental results on three real-world datasets demonstrate that the proposed model significantly outperforms existing state-of-the-art baseline models in terms of performance.}, archive = {J_DKE}, author = {Yuanming Zhang and Ziyou He and Yongbiao Lou and Haixia Long and Fei Gao}, doi = {10.1016/j.datak.2025.102522}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102522}, shortjournal = {Data Knowl. Eng.}, title = {Temporal knowledge graph recommendation with sequence-aware and path reasoning}, volume = {161}, year = {2026}, }
• (2026). Secure data storage in multi-cloud environments using lattice-based saber with diffie-hellman cryptography and authenticate based on PUF-ECC. DKE, 161, 102512. (www)

  @article{ , abstract = {Human life has become highly dependent on data in recent decades almost every facet of daily activities, leading to its storage in multi-cloud environments. To ensure data integrity, confidentiality, and privacy, it is essential to protect data from unauthorized access. This paper proposes a novel approach for securing data in multi-cloud environments for user authentication and data storage using Lattice-Based Saber Cryptography combined with PUF-ECC and the Enhanced Goose Optimization Algorithm (EGOA). The initial user authentication is achieved through the PUF-ECC digital signature algorithm, which verifies both the user's and the device's identity. Once authenticated, user data is securely transmitted to the cloud server based on Lattice-Based Saber post-quantum cryptography combined with the Diffie-Hellman key exchange protocol. The encrypted data is then stored across multiple cloud storage through a cloud controller using RAM-based chunking. For efficient data retrieval, the Enhanced Goose Optimization Algorithm (EGOA) is employed to extract encrypted data from clouds. Finally, the data is decrypted using the Lattice-Based Saber decryption algorithm and securely retrieved by the authenticated user. This method enhances both the security and efficiency of cloud data management and retrieval. The experiment is carried out with the proposed methodologies and also compared with the existing technologies. The proposed approach achieves encryption times of 9.68 ms, key generation times of 4.84 ms, and block creation times of 1.59 ms, while maintaining a 93.7 % confidentiality rate, a 98 % packet delivery ratio, a transmission delay of 0.026 ms, throughput of 407.33 MB/s, jitter of 3.26 ms, and an RTT of 0.17 ms, demonstrating its effectiveness in secure data storage and retrieval in multi-cloud environments.}, archive = {J_DKE}, author = {R. Iyswarya and R. Anitha}, doi = {10.1016/j.datak.2025.102512}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102512}, shortjournal = {Data Knowl. Eng.}, title = {Secure data storage in multi-cloud environments using lattice-based saber with diffie-hellman cryptography and authenticate based on PUF-ECC}, volume = {161}, year = {2026}, }
• (2026). ASF: A novel associative scoring function for embedded knowledge graph reasoning. DKE, 161, 102511. (www)

  @article{ , abstract = {One of the most important tools for knowledge management is the Knowledge Graph (KG), a multi-relational graph that depicts rich factual information across entities. A KG represents entities as nodes and relations as edges, with each edge represented by a triplet: (head entity, relation, tail entity). The Scoring Function (SF) in a KG quantifies the plausibility of these triplets and is often derived from KG embeddings. However, due to the distinct relational patterns across KGs, an SF that performs well on one KG might fail on another, making the design of optimal SFs a challenging task. This study introduces the concept of an Associative Scoring Function (ASF), which leverages Association Rule Mining (ARM) to discover and incorporate patterns and characteristics of symmetric, asymmetric, inverse, and other relational types within embedded KGs. The ARM technique in ASF uses the FP-Growth algorithm to extract meaningful associations, which is enhanced further through hyperparameter tuning. Extensive experiments on benchmark datasets demonstrate that ASF is KG-independent and performs better than state-of-the-art SFs. These results highlight ASF's potential to generalize across diverse KGs, offering a significant advancement in the KG link prediction task.}, archive = {J_DKE}, author = {MVPT Lakshika and HA Caldera}, doi = {10.1016/j.datak.2025.102511}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102511}, shortjournal = {Data Knowl. Eng.}, title = {ASF: A novel associative scoring function for embedded knowledge graph reasoning}, volume = {161}, year = {2026}, }
• (2026). A graph-based model for semantic textual similarity measurement. DKE, 161, 102509. (www)

  @article{ , abstract = {Measuring semantic similarity between sentence pairs is a fundamental problem in Natural Language Processing with applications in various domains, including machine translation, speech recognition, automatic question answering, and text summarization. Despite its significance, accurately assessing semantic similarity remains a challenging task, particularly for underrepresented languages such as Vietnamese. Existing methods have yet to fully leverage the unique linguistic characteristics of Vietnamese for semantic similarity measurement. To address this limitation, we propose GBNet-STS (Graph-Based Network for Semantic Textual Similarity), a novel framework for measuring the semantic similarity of Vietnamese sentence pairs. GBNet-STS integrates lexical-grammatical similarity scores and distributional semantic similarity scores within a multi-layered graph-based model. By capturing different semantic perspectives through multiple interconnected layers, our approach provides a more comprehensive and robust similarity estimation. Experimental results demonstrate that GBNet-STS outperforms traditional methods, achieving state-of-the-art performance in Vietnamese semantic similarity tasks.}, archive = {J_DKE}, author = {Van-Tan Bui and Quang-Minh Nguyen and Van-Vinh Nguyen and Duc-Toan Nguyen}, doi = {10.1016/j.datak.2025.102509}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102509}, shortjournal = {Data Knowl. Eng.}, title = {A graph-based model for semantic textual similarity measurement}, volume = {161}, year = {2026}, }
• (2026). Rule-guided process discovery. DKE, 161, 102508. (www)

  @article{ , abstract = {Event data extracted from information systems serves as the foundation for process mining, enabling the extraction of insights and identification of improvements. Process discovery focuses on deriving descriptive process models from event logs, which form the basis for conformance checking, performance analysis, and other applications. Traditional process discovery techniques predominantly rely on event logs, often overlooking supplementary information such as domain knowledge and process rules. These rules, which define relationships between activities, can be obtained through automated techniques like declarative process discovery or provided by domain experts based on process specifications. When used as an additional input alongside event logs, such rules have significant potential to guide process discovery. However, leveraging rules to discover high-quality imperative process models, such as BPMN models and Petri nets, remains an underexplored area in the literature. To address this gap, we propose an enhanced framework, IMr, which integrates discovered or user-defined rules into the process discovery workflow via a novel recursive approach. The IMr framework employs a divide-and-conquer strategy, using rules to guide the selection of process structures at each recursion step in combination with the input event log. We evaluate our approach on several real-world event logs and demonstrate that the discovered models better align with the provided rules without compromising their conformance to the event log. Additionally, we show that high-quality rules can improve model quality across well-known conformance metrics. This work highlights the importance of integrating domain knowledge into process discovery, enhancing the quality, interpretability, and applicability of the resulting process models.}, archive = {J_DKE}, author = {Ali Norouzifar and Marcus Dees and Wil van der Aalst}, doi = {10.1016/j.datak.2025.102508}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102508}, shortjournal = {Data Knowl. Eng.}, title = {Rule-guided process discovery}, volume = {161}, year = {2026}, }
• (2026). LQ-FJS: A logical query-digging fake-news judgment system with structured video-summarization engine using LLM. DKE, 161, 102507. (www)

  @article{ , abstract = {The proliferation of online social platforms can greatly benefit people by fostering remote relationships, but it also inevitably amplifies the impact of multimodal fake news on societal trust and ethics. Existing fake-news detection AI systems are still vulnerable to the inconspicuous and indiscernible multimodal misinformation, and often lacking interpretability and accuracy in cross-platform settings. Hence, we propose a new innovative logical query-digging fake-news judgment system (LQ-FJS) to tackle the above problem based on multimodal approach. The LQ-FJS verifies the truthfulness of claims made within multimedia news by converting video content into structured textual summaries. It then acts as an interpretable agent, explaining the reasons for identified fake news by the structured video-summarization engine (SVSE) to act as an interpretable detection intermediary agent. The SVSE generates condensed captions for raw video content, converting it into structured textual narratives. Then, LQ-FJS exploits these condensed captions to retrieve reliable information related to the video content from LLM. Thus, LQ-FJS cross-verifies external knowledge sources and internal LLM responses to determine whether contradictions exist with factual information through a multimodal inconsistency verification procedure. Our experiments demonstrate that the subtle summarization produced by SVSE can facilitate the generation of explanatory reports that mitigate large-scale trust deficits caused by opaque “black-box” models. Our experiments show that LQ-FJS improves F1 scores by 4.5% and 7.2% compared to state-of-the-art models (FactLLaMA 2023 and HiSS 2023), and increases 14% user trusts through interpretable conclusions.}, archive = {J_DKE}, author = {Jhing-Fa Wang and Din-Yuen Chan and Hsin-Chun Tsai and Bo-Xuan Fang}, doi = {10.1016/j.datak.2025.102507}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102507}, shortjournal = {Data Knowl. Eng.}, title = {LQ-FJS: A logical query-digging fake-news judgment system with structured video-summarization engine using LLM}, volume = {161}, year = {2026}, }
• (2026). ABBA: Index structure for sequential pattern-based aggregate queries. DKE, 161, 102506. (www)

  @article{ , abstract = {Pattern-based aggregate (PBA) queries constitute an important and widely used type of analytical queries in sequence OLAP (S-OLAP) systems. Unfortunately, finding accurate answers to PBA queries in the S-OLAP system is often very expensive both in terms of time and memory consumption. In this paper we propose an efficient and easily maintainable index structure called the ABBA Index, which addresses the problem of PBA query processing. Experiments conducted using the KDD Cup data and public transport passengers’ travel behavior data show that our index outperforms state-of-the art solutions while requiring much less memory. The ABBA Index can be easily extended to support pattern-based aggregate queries over hierarchy (PBA-H), a novel class of analytical queries which we introduce as the second main contribution of the paper. Sensitivity, scalability and complexity analysis of the ABBA Index is also provided.}, archive = {J_DKE}, author = {Witold Andrzejewski and Tadeusz Morzy and Maciej Zakrzewicz}, doi = {10.1016/j.datak.2025.102506}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102506}, shortjournal = {Data Knowl. Eng.}, title = {ABBA: Index structure for sequential pattern-based aggregate queries}, volume = {161}, year = {2026}, }
• (2026). Elevating human-machine collaboration in NLP for enhanced content creation and decision support. DKE, 161, 102505. (www)

  @article{ , abstract = {Human-machine collaboration in Natural Language Processing (NLP) is revolutionizing content creation and decision support by seamlessly combining the strengths of both entities for enhanced efficiency and quality. The lack of seamless integration between human creativity and machine efficiency in NLP hinders optimal content creation and decision support. The objective of this study is to explore and promote the integration of human-machine collaboration in NLP to enhance both content creation and decision support processes. Data Acquisition for NLP requests involves defining the task and target audience, identifying relevant data sources like text documents and web data, and incorporating human expertise for data curation through validation and annotation. Machine processing techniques like tokenization, stemming/lemmatization, and removal of stop words, as well as human input for tasks like data annotation and error correction, to improve data quality and relevance for NLP applications. The combination of automated processing and human feedback leads to more precise and dependable effects. Techniques such as sentiment analysis, topic modelling, and entity recognition are utilized to excerpt valued perceptions from the data and enhance collaboration between humans and machines. These techniques help to streamline the NLP process and ensure that the system is providing accurate and relevant information to users. The analysis of NLP models in machine processing involves training the models to perform specific tasks, such as summarization, sentiment analysis, information extraction, trend identification, and creative content generation. The results show that social media leads with 90% usage, pivotal for audience engagement, while blogs at 78% highlight their depth in content creation implementation using Python software. These trained models are then used to improve decision-making processes, generate creative content, and enhance the accuracy of search results. The future scope involves leveraging advanced NLP techniques to deepen the collaboration between humans and machines for more effective content creation and decision support.}, archive = {J_DKE}, author = {Priyanka V. Deshmukh and Aniket K. Shahade}, doi = {10.1016/j.datak.2025.102505}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102505}, shortjournal = {Data Knowl. Eng.}, title = {Elevating human-machine collaboration in NLP for enhanced content creation and decision support}, volume = {161}, year = {2026}, }
• (2026). ELEVATE-ID: Extending large language models for end-to-end entity linking evaluation in indonesian. DKE, 161, 102504. (www)

  @article{ , abstract = {Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of natural language processing tasks. However, their effectiveness in low-resource languages remains underexplored, particularly in complex tasks such as end-to-end Entity Linking (EL), which requires both mention detection and disambiguation against a knowledge base (KB). In earlier work, we introduced IndEL — the first end-to-end EL benchmark dataset for the Indonesian language — covering both a general domain (news) and a specific domain (religious text from the Indonesian translation of the Quran), and evaluated four traditional end-to-end EL systems on this dataset. In this study, we propose ELEVATE-ID, a comprehensive evaluation framework for assessing LLM performance on end-to-end EL in Indonesian. The framework evaluates LLMs under both zero-shot and fine-tuned conditions, using multilingual and Indonesian monolingual models, with Wikidata as the target KB. Our experiments include performance benchmarking, generalization analysis across domains, and systematic error analysis. Results show that GPT-4 and GPT-3.5 achieve the highest accuracy in zero-shot and fine-tuned settings, respectively. However, even fine-tuned GPT-3.5 underperforms compared to DBpedia Spotlight — the weakest of the traditional model baselines — in the general domain. Interestingly, GPT-3.5 outperforms Babelfy in the specific domain. Generalization analysis indicates that fine-tuned GPT-3.5 adapts more effectively to cross-domain and mixed-domain scenarios. Error analysis uncovers persistent challenges that hinder LLM performance: difficulties with non-complete mentions, acronym disambiguation, and full-name recognition in formal contexts. These issues point to limitations in mention boundary detection and contextual grounding. Indonesian-pretrained LLMs, Komodo and Merak, reveal core weaknesses: template leakage and entity hallucination, respectively—underscoring architectural and training limitations in low-resource end-to-end EL. 1}, archive = {J_DKE}, author = {Ria Hari Gusmita and Asep Fajar Firmansyah and Hamada M. Zahera and Axel-Cyrille Ngonga Ngomo}, doi = {10.1016/j.datak.2025.102504}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102504}, shortjournal = {Data Knowl. Eng.}, title = {ELEVATE-ID: Extending large language models for end-to-end entity linking evaluation in indonesian}, volume = {161}, year = {2026}, }
• (2026). Time-aware complex question answering over temporal knowledge graph. DKE, 161, 102503. (www)

  @article{ , abstract = {Knowledge Graph Question Answering (KGQA) is a crucial topic in Knowledge Graphs (KGs), with the objective of retrieving the corresponding facts from KGs to answer given questions. In practical applications, facts in KGs usually have time constraints, thus, question answering on Temporal Knowledge Graphs (TKGs) has attracted extensive attention. Existing Temporal Knowledge Graph Question Answering (TKGQA) methods focus on dealing with complex questions involving multiple facts, and mainly face two challenges. First, these methods only consider matching questions with facts in TKGs to identify the answer, ignoring the temporal order between different facts, which makes it challenging to solve the questions involving temporal order. Second, they usually focus on the representation of the question text while neglecting the rich semantic information within the questions, which leads to certain limitations in understanding question. To address the above challenges, this research proposes a model named Time-Aware Complex Question Answering (TA-CQA). Specifically, we extend the Temporal Knowledge Graph Embedding (TKGE) model by incorporating temporal order information into the embedding vectors, ensuring that the model can distinguish the temporal order of different facts. To enhance the semantic representation of the question, we integrate question information using attention mechanism and learnable encoder. Different from the previous TKGQA methods, we propose time relevance measurement to further enhance the accuracy of answer prediction by better capturing the correlation between question information and time information. Multiple sets of experiments on CronQuestions and TimeQuestions demonstrate our model’s superior performance across all question types. In particular, for complex questions involving multiple facts, the hit@1 values are increased by 3.2% and 3.5% respectively.}, archive = {J_DKE}, author = {Luyi Bai and Tongyue Zhang and Guangchen Feng}, doi = {10.1016/j.datak.2025.102503}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102503}, shortjournal = {Data Knowl. Eng.}, title = {Time-aware complex question answering over temporal knowledge graph}, volume = {161}, year = {2026}, }
• (2026). Conceptual modeling of user perspectives — From data warehouses to alliance-driven data ecosystems. DKE, 161, 102502. (www)

  @article{ , abstract = {The increasing complexity of modern information systems has highlighted the need for advanced conceptual modeling techniques that incorporate multi-perspective and view-based approaches. This paper explores the role of multi-perspective modeling and view modeling in designing distributed, heterogeneous systems while addressing diverse user requirements and ensuring semantic consistency. These methods enable the representation of multiple viewpoints, traceability, and dynamic integration across different levels of abstraction. Key advancements in schema mapping, view maintenance, and semantic metadata management are examined, illustrating how they support query optimization, data quality, and interoperability. We discuss how data management architectures, such as data ecosystems, data warehouses, and data lakes, leverage these innovations to enable flexible and sustainable data sharing. By integrating user-centric and goal-oriented modeling frameworks, the alignment of technical design with organizational and social requirements is emphasized. Future challenges include the need for enhanced reasoning capabilities and collaborative tools to manage the growing complexity of interconnected systems while maintaining adaptability and trust.}, archive = {J_DKE}, author = {Sandra Geisler and Christoph Quix and István Koren and Matthias Jarke}, doi = {10.1016/j.datak.2025.102502}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102502}, shortjournal = {Data Knowl. Eng.}, title = {Conceptual modeling of user perspectives — From data warehouses to alliance-driven data ecosystems}, volume = {161}, year = {2026}, }
• (2026). Source-free domain adaptation with complex distribution considerations for time series data. DKE, 161, 102501. (www)

  @article{ , abstract = {Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained model from a labeled source domain to an unlabeled target domain without accessing source domain data, thereby protecting source domain privacy. Although SFDA has recently been applied to time series data, the inherent complex distribution characteristics including temporal variability and distributional diversity of such data remain underexplored. Time series data exhibit significant dynamic variability influenced by collection environments, leading to discrepancies between sequences. Additionally, multidimensional time series data face distributional diversity across dimensions. These complex characteristics increase the learning difficulty for source models and widen the adaptation gap between the source and target domains. To address these challenges, this paper proposes a novel SFDA method for time series data, named Adaptive Latent Subdomain feature extraction and joint Prediction (ALSP). The method divides the source domain, which has a complex distribution, into multiple latent subdomains with relatively simple distributions, thereby effectively capturing the features of different subdistributions. It extracts latent domain-specific and domain-invariant features to identify subdomain-specific characteristics. Furthermore, it combines domain-specific classifiers and a domain-invariant classifier to enhance model performance through multi-classifier joint prediction. During target domain adaptation, ALSP reduces domain dependence by extracting invariant features, thereby narrowing the distributional gap between the source and target domains. Simultaneously, it leverages prior knowledge from the source domain distribution to support the hypothesis space and dynamically adapt to the target domain. Experiments on three real-world datasets demonstrate that ALSP achieves superior performance in cross-domain time series classification tasks, significantly outperforming existing methods.}, archive = {J_DKE}, author = {Jing Shang and Zunming Chen and Zhiwen Xiao and Zhihui Wu and Yifei Zhang and Jibing Wang}, doi = {10.1016/j.datak.2025.102501}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102501}, shortjournal = {Data Knowl. Eng.}, title = {Source-free domain adaptation with complex distribution considerations for time series data}, volume = {161}, year = {2026}, }
• (2026). Conceptual modeling: A large language model assistant for characterizing research contributions. DKE, 161, 102497. (www)

  @article{ , abstract = {The body of conceptual modeling research publications is vast and diverse, making it challenging for a single researcher or research group to fully comprehend the field’s overall development. Although some approaches have been proposed to help organize these research contributions, it is still unrealistic to expect human experts to manually comprehend and characterize all of this research. However, as generative AI tools based on large language models, such as ChatGPT, become increasingly sophisticated, it may be possible to replace or augment tedious, manual work with semi-automated approaches. In this research, we present a customized version of ChatGPT that is tuned to the task of characterizing conceptual modeling research. Experiments with this AI tool demonstrate that it is feasible to create a usable knowledge survey for the continually evolving body of conceptual modeling research contributions.}, archive = {J_DKE}, author = {Stephen W. Liddle and Heinrich C. Mayr and Oscar Pastor and Veda C. Storey and Bernhard Thalheim}, doi = {10.1016/j.datak.2025.102497}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102497}, shortjournal = {Data Knowl. Eng.}, title = {Conceptual modeling: A large language model assistant for characterizing research contributions}, volume = {161}, year = {2026}, }
• (2026). Dynamic time warping for classifying long-term trends in time series. DKE, 161, 102495. (www)

  @article{ , abstract = {This paper explores the potential of dynamic time warping (DTW) for recognizing different segments in time series data characterized by their long-term trends and curvature. To perform classification, a set of reference data for each class is required, where each time series in the reference set represents a typical shape of that class. The classification process involves computing the DTW distance between a given time series and each reference time series, then assigning the time series to the class with the minimum distance. Experiments on both simulated and real-world time series data from two different use cases demonstrate that DTW can correctly classify the different segments. Additionally, the paper investigates whether incorrectly classified phases could indicate data security issues. Additional experiments are performed to assess the number of data points required to reliably classify a segment correctly. These experiments highlight the limitations and emphasize the importance of selecting good reference data.}, archive = {J_DKE}, author = {Anna-Christina Glock and Klaus Chmelina and Johannes Fürnkranz and Thomas Hütter}, doi = {10.1016/j.datak.2025.102495}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102495}, shortjournal = {Data Knowl. Eng.}, title = {Dynamic time warping for classifying long-term trends in time series}, volume = {161}, year = {2026}, }
• (2026). Semantic-aware query answering with large language models. DKE, 161, 102494. (www)

  @article{ , abstract = {In the modern data-driven world, answering queries over heterogeneous and semantically inconsistent data remains a significant challenge. Modern datasets originate from diverse sources, such as relational databases, semi-structured repositories, and unstructured documents, leading to substantial variability in schemas, terminologies, and data formats. Traditional systems, constrained by rigid syntactic matching and strict data binding, struggle to capture critical semantic connections and schema ambiguities, failing to meet the growing demand among data scientists for advanced forms of flexibility and context-awareness in query answering. In parallel, the advent of Large Language Models (LLMs) has introduced new capabilities in natural language interpretation, making them highly promising for addressing such challenges. However, LLMs alone lack the systematic rigor and explainability required for robust query processing and decision-making in high-stakes domains. In this paper, we propose Soft Query Answering (Soft QA), a novel hybrid approach that integrates LLMs as an intermediate semantic layer within the query processing pipeline. Soft QA enhances query answering adaptability and flexibility by injecting semantic understanding through context-aware, schema-informed prompts, and leverages LLMs to semantically link entities, resolve ambiguities, and deliver accurate query results in complex settings. We demonstrate its practical effectiveness through real-world examples, highlighting its ability to resolve semantic mismatches and improve query outcomes without requiring extensive data cleaning or restructuring.}, archive = {J_DKE}, author = {Paolo Atzeni and Teodoro Baldazzi and Luigi Bellomarini and Eleonora Laurenza and Emanuel Sallinger}, doi = {10.1016/j.datak.2025.102494}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102494}, shortjournal = {Data Knowl. Eng.}, title = {Semantic-aware query answering with large language models}, volume = {161}, year = {2026}, }
• (2026). An integrated requirements framework for analytical and AI projects. DKE, 161, 102493. (www)

  @article{ , abstract = {To this day, the requirements of data warehouses, user visualizations and ML projects have been tackled in an independent manner, ignoring the possible cross-requirements, collective constraints and dependencies between the outputs of the different systems that should be taken into account to ensure a successful analytical project. In this work, we take a holistic approach and propose a methodology that supports modeling and subsequent analysis while taking into account these three aspects. This methodology has several advantages, mainly that (i) it enables us to identify possible conflicts between actors on different tasks that are overlooked if the systems are treated in an isolated manner and (ii) this holistic view enables modeling multi-company systems, where the information or even the analytical results can be provided by third-parties, identifying key participants in federated environments. After presenting the required formalism to carry out this kind of analysis, we showcase it on a real-world running example of the tourism sector.}, archive = {J_DKE}, author = {Juan Trujillo and Ana Lavalle and Alejandro Reina-Reina and Jorge García-Carrasco and Alejandro Maté and Wolfgang Maaß}, doi = {10.1016/j.datak.2025.102493}, journal = {Data & Knowledge Engineering}, month = {1}, pages = {102493}, shortjournal = {Data Knowl. Eng.}, title = {An integrated requirements framework for analytical and AI projects}, volume = {161}, year = {2026}, }

DSS - 1

• (2025). AI art in the gig economy: Investigating the effects of non-copyrightability in online labor markets. DSS, 199, 114545. (www)

  @article{ , abstract = {As generative AI continues to transform industries, including the creative sector, it has become critical to understand how it interacts with legal frameworks. This study aims to investigate the effect of the landmark ruling issued by the U.S. District Court on August 18, 2023, which declared AI-generated art uncopyrightable to provide clarity to previously ambiguous legal standards on the AI-related services in online labor markets. Our findings reveal that prices for AI-related gigs on an online freelancer platform dropped by 32.97 % following the ruling, suggesting that the lack of copyright may have reduced the perceived value by limiting clients' residual rights. Furthermore, our research indicates that both freelancer experience and communication efficiency significantly moderate the relationship between AI art non-copyrightability and project pricing. In addition, the results show that large corporate clients were more affected by the ruling than individual clients. In contrast, prices for projects commissioned by small and mid-sized corporate clients did not change significantly. This suggests that large firms are more sensitive to intellectual property uncertainties because they rely heavily on formal rights to secure control and revenue from creative assets. This research contributes to a nuanced understanding of how legal frameworks for AI may shape the gig economy's AI art-related creative services, offering valuable guidelines for more informed decision-making by freelancers, clients, platform owners, and policymakers in this evolving landscape.}, archive = {J_DSS}, author = {Lan Li and Noelle Li Ying Cheah and Seung Hyun Kim}, doi = {10.1016/j.dss.2025.114545}, journal = {Decision Support Systems}, month = {12}, pages = {114545}, shortjournal = {Decis. Supp. Syst.}, title = {AI art in the gig economy: Investigating the effects of non-copyrightability in online labor markets}, volume = {199}, year = {2025}, }

EAAI - 512

• (2025). Teach sample-specific knowledge: Separated distillation based on samples. EAAI, 162, 112696. (www)

  @article{ , abstract = {Recent advancements in deep neural networks have revolutionized computer vision, enabling practical applications like classification and object detection. However, deploying these models on resource-constrained devices remains a critical challenge due to their high computational demands. Knowledge Distillation (KD) has emerged as an effective technique to address this issue by transferring knowledge from complex teacher models to lightweight student models, enhancing efficiency while maintaining high performance. Traditional logit-based KD methods use forward Kullback–Leibler divergence (FKLD) to transfer meaningful knowledge. However, FKLD typically exhibits a mode-averaging property, causing students to focus on non-target information, whether the teacher’s samples are correct or incorrect. Additionally, when handling uncertain samples, even teacher models may fail to classify them accurately, leading to incorrect predictions and confusing the students. To address these issues, we classify the dataset into two groups based on the teacher’s predictions: correct and incorrect samples. To ensure a more reliable transfer of knowledge from teacher to student for correct samples, we employ both forward Kullback–Leibler divergence (FKLD) and reverse Kullback–Leibler divergence (RKLD), which has mode-focusing properties. We also reduce temperature scaling for RKLD to enhance the focus on target information, ensuring that the student model prioritizes meaningful knowledge while minimizing the influence of non-target information. Conversely, for incorrect predictions, our method minimizes the teacher’s knowledge, encouraging students to rely more on the true labels by focusing on cross-entropy loss. Experimental results on both classification and object detection tasks demonstrate that our method, Teach Sample-Specific Knowledge (TSSK), outperforms state-of-the-art KD methods, making it ideal for deployment on-devices in real-world scenarios.}, archive = {J_EAAI}, author = {Seonghak Kim and Gyeongdo Ham and Suin Lee and Daeshik Kim}, doi = {10.1016/j.engappai.2025.112696}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112696}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Teach sample-specific knowledge: Separated distillation based on samples}, volume = {162}, year = {2025}, }
• (2025). A context-adaptive and dual-dimensional collaboration method for infrared-visible image registration and fusion. EAAI, 162, 112684. (www)

  @article{ , abstract = {Infrared and visible image registration and fusion are tightly coupled tasks. However, in unregistered scenarios, prioritizing registration accuracy can compromise the integrity of complementary information, whereas enforcing fusion quality may disrupt geometric consistency. To address this, we propose a Context-Adaptive and Dual-Dimensional Collaboration Method for Infrared-Visible Image Registration and Fusion. Firstly, we design an adaptive context-aware channel fusion convolution module that simultaneously serves feature extraction for both registration and fusion tasks, perceiving contextual feature relationships to coordinate corresponding regions across modalities while dynamically generating task-adaptive feature representations. Secondly, we design a dual-dimensional dynamic collaboration module leveraging interactive mapping across spatial and channel dimensions to effectively fuse complementary cross-modal features while preserving detail and texture information. Finally, the registered and fused images are reconstructed by the spatial transformer network and lightweight decoder. Extensive tests on multiple datasets show that our method exhibits excellent performance in visual effects, quantitative metrics, and generalization ability, making it suitable for application scenarios such as security monitoring, fault detection, and target recognition.}, archive = {J_EAAI}, author = {Minghao Jiang and Shaoshu Gao and Xiaodong Zhang and Xingli Wang and Qing Hu and Weiming Wang}, doi = {10.1016/j.engappai.2025.112684}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112684}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A context-adaptive and dual-dimensional collaboration method for infrared-visible image registration and fusion}, volume = {162}, year = {2025}, }
• (2025). A lightweight vision mamba coding UNet for medical image segmentation. EAAI, 162, 112676. (www)

  @article{ , abstract = {The segmentation of medical images is vital for advancing research in medicine and supporting precise clinical diagnoses. Over the past few years, neural network-based methods have become a major focus of research and have been broadly adopted in medical image segmentation. However, mainstream methods like Transformer have excessive computational cost requirements, which makes them impractical for mobile medical applications. Therefore, this paper proposes a Lightweight Vision Mamba Coding UNet (LVMC-UNet), which integrates the Rotation-based Vision Mamba module (RVM) and Correlation Space Fusion module (CSF) in a lightweight manner. The RVM module processes the input images in parallel and introduces rotary positional encoding (RPE) on the Vision Mamba architecture to address the limitation of the Mamba module’s insufficient ability to capture local details. The CSF module integrates multi-stage and multi-scale feature maps and introduces blueprint separable convolutions (BSConvs) to enhance intra-kernel correlation. Comprehensive experiments on three datasets demonstrate that LVMC-UNet outperforms current lightweight segmentation methods and strikes a balance between segmentation accuracy and computational requirements.}, archive = {J_EAAI}, author = {Yuanyuan Li and Yifei Duan and Guanqiu Qi and Baisen Cong and Li Zhang and Zhiqin Zhu}, doi = {10.1016/j.engappai.2025.112676}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112676}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight vision mamba coding UNet for medical image segmentation}, volume = {162}, year = {2025}, }
• (2025). A real-time design and implementation of intelligent drowsiness and fatigue recognition system for enhancing driver safety. EAAI, 162, 112665. (www)

  @article{ , abstract = {In recent years, the rise in personal vehicle usage for speed and convenience has increased road accidents, with driver fatigue as a major cause. Existing drowsiness detection systems suffer from high computational complexity and poor generalization across diverse conditions. To address these challenges, we propose an Intelligent Drowsiness and Fatigue Recognition (IDFR) System leveraging deep learning for real-time driver monitoring. Our system employs a customized Convolutional Neural Network (CNN) optimized for detecting eye states (open or closed) with high accuracy. A key feature is the use of the Eye Aspect Ratio (EAR) with a threshold of 0.2, empirically validated for precise drowsiness detection across varying lighting conditions and facial features. The system is trained on the Media Research Lab (MRL) Eye dataset, containing over 84,000 images, ensuring robust generalization across diverse demographics. The proposed IDFR system achieves 98.50 % accuracy, outperforming state-of-the-art deep learning models including Visual Geometry Group 16-layer network (VGG16), Visual Geometry Group 19-layer network (VGG19), Residual Network 152-layer (ResNet152), and Densely Connected Convolutional Network 201-layer (DenseNet201) in accuracy, precision, recall, and computational efficiency. By integrating transfer learning and fine-tuning techniques, we significantly reduce computational costs while maintaining high performance. The system was evaluated under challenging conditions, such as low lighting, high glare, and partial occlusion (e.g., sunglasses, shadows, and face coverings), ensuring robust detection accuracy. Our results demonstrate the IDFR system's potential for real-time in-vehicle deployment, providing immediate driver feedback and reducing accident risks. This work advances intelligent driver assistance technologies, enhancing road safety and future transportation systems.}, archive = {J_EAAI}, author = {Samy Abd El-Nabi and Khalil F. Ramadan and El-Sayed M. El-Rabaie and Ahmed Emam and Walid El-Shafai}, doi = {10.1016/j.engappai.2025.112665}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112665}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A real-time design and implementation of intelligent drowsiness and fatigue recognition system for enhancing driver safety}, volume = {162}, year = {2025}, }
• (2025). Coupled fault diagnosis for centrifugal pumps through boruta-shap feature selection and rime-enhanced stacked denoised autoencoder. EAAI, 162, 112659. (www)

  @article{ , abstract = {This study addresses the critical engineering challenges of diagnosing multi-condition cavitation and impeller erosion damage in deep sea centrifugal pumps, where traditional methods suffer from feature redundancy and poor generalization. An artificial intelligence (AI)-driven framework is proposed by integrating Boruta-Shap interpretable feature selection with a Rime-optimized Stacked Denoising Auto-Encoder (SDAE). Firstly, vibration signals undergo multi-domain feature extraction. The Boruta-Shap method quantifies feature interactions via Shapley values and shadow feature hypothesis testing, enabling physics-informed selection of cavitation energy indicators and impeller damage transient features, reducing feature dimensionality by approximately 70 %. Secondly, the Rime algorithm, inspired by ice-crystal growth dynamics, globally optimizes SDAE hyperparameters via balancing soft-rime exploration and hard-rime exploitation. Validated experimentally under variable conditions (0.8, 1.0 and 1.2 Q d ), the framework achieves 96 % testing accuracy, outperforming other diagnostic models, with reduced cross-condition misclassification. By bridging interpretable AI with multi-physics diagnostics, this work provides a replicable solution for predictive maintenance of energy infrastructure, advancing intelligent maintenance technologies for industrial systems facing complex hydrodynamic perturbations.}, archive = {J_EAAI}, author = {Kang Hu and Hui Sun and Wei Fan and Qiaorui Si and Yu Wu and Shouqi Yuan}, doi = {10.1016/j.engappai.2025.112659}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112659}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Coupled fault diagnosis for centrifugal pumps through boruta-shap feature selection and rime-enhanced stacked denoised autoencoder}, volume = {162}, year = {2025}, }
• (2025). Multimodal interval prediction of carbon dioxide emissions from heavy construction machinery: A missing-data robust inverted-transformer model considering sensors failure in complex construction environment. EAAI, 162, 112658. (www)

  @article{ , abstract = {Accurate forecasting of carbon dioxide (CO 2 ) emissions from heavy construction machinery in large-scale infrastructure projects presents a viable avenue for mitigating climate change. However, most current studies neglect the intricate operating conditions of such machinery. Moreover, in complex construction environments with high-frequency vibrations and heavy dust, sensor failures leading to random data loss significantly increase the difficulty of emission prediction. Especially when high-emission periods account for a small proportion, existing methods struggle to effectively capture CO 2 emission peaks. To address these challenges, this study proposes an improved Inverted-Transformer (iTransformer) multimodal interval prediction model for accurate CO 2 emission forecasting in heavy construction machinery under conditions of random sensor failures. The model incorporates a Mixture of Experts (MoE) mechanism within the iTransformer framework, which adaptively adjusts expert weights for missing modalities, thereby reducing prediction errors caused by random data loss. Additionally, to capture localized CO 2 emission peaks, this study introduces a Peak Capture Loss (PCL) function, which adjusts incremental emissions between adjacent time steps by supervising the differences between generated sequences, enabling the model to track abrupt emission variations. The Bootstrap method is also utilised to quantify and estimate uncertainty in the CO 2 emission Interval Prediction. Case studies reveal that the proposed model achieves high prediction accuracy (coefficient of determination ( R 2 ) = 0.99), especially across various data missing rates (5 %, 10 %, 15 %), with the average R 2 value increasing by approximately 6 %. This provides a novel approach for predicting emissions of heavy construction machinery in large-scale infrastructure projects.}, archive = {J_EAAI}, author = {Zhouquan Dong and Xiaoling Wang and Jun Zhang and Peng Yu and Zhijian Cai}, doi = {10.1016/j.engappai.2025.112658}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112658}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multimodal interval prediction of carbon dioxide emissions from heavy construction machinery: A missing-data robust inverted-transformer model considering sensors failure in complex construction environment}, volume = {162}, year = {2025}, }
• (2025). Enhancing wind power forecasting accuracy under extreme weather: Leveraging a dual-model approach with condition-based classification. EAAI, 162, 112656. (www)

  @article{ , abstract = {Accurate wind power forecasting is essential for grid stability, yet extreme weather events, exacerbated by climate change—pose challenges that current methods often overlook. In this study, we introduce a forecasting framework that accounts for extreme weather conditions to improve forecasting accuracy. First, we correct Numerical Weather Prediction (NWP) via temporal alignment and bias correction. Then, we address the differences in data distributions between typical and extreme weather conditions, while accounting for multivariate meteorological influences. We establish weather classification thresholds by integrating meteorological indicators (wind speed variation rate, precipitation, snowfall) with a data-driven sensitivity analysis, and optimize them via grid search to maximize the coefficient of determination ( R 2 ) while balancing false alarms and detection rates. Finally, due to the difference in distribution and data pattern, we implemented distinct forecasting models for each condition. A Convolutional Neural Network (CNN) models normal conditions by capturing temporal dependencies, while Light Gradient Boosting Machine (LightGBM) handles sparse extreme-weather samples. Compared to models that ignore extreme weather, our approach reduces Mean Absolute Error (MAE) by 21.13% and Root Mean Squared Error (RMSE) by 9.29%. Under extreme conditions, Mean Absolute Percentage Error (MAPE) is reduced by 32.76%, demonstrating enhanced robustness in handling extreme weather scenarios.}, archive = {J_EAAI}, author = {Weimin Yuan and Han Yang and Zhu Han and Yanru Zhang}, doi = {10.1016/j.engappai.2025.112656}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112656}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing wind power forecasting accuracy under extreme weather: Leveraging a dual-model approach with condition-based classification}, volume = {162}, year = {2025}, }
• (2025). Integrating multimodal biophysical features with hybrid deep learning for ribonucleic acid secondary structure prediction. EAAI, 162, 112648. (www)

  @article{ , abstract = {The secondary structure of ribonucleic acid (RNA) is pivotal for elucidating its functional roles. However, existing deep learning models predominantly rely on single-feature representations, which restricts their capacity to sufficiently capture the intricate information embedded in RNA sequences. To solve this problem, we propose a novel RNA secondary structure prediction method based on multimodal feature fusion and hybrid deep learning. By integrating multiple features of RNA are modeled to achieve the synergistic effect of the chemical microenvironment, local physical constraints and long-range interactions. We conceptualize RNA structure as an image and treat various RNA features as distinct channels of that image. To facilitate the transformation from RNA sequences to RNA images, we design a hybrid neural network architecture. Specifically, the sequence feature extraction module first extracts features directly from the RNA sequence. These features are then passed to the image feature extraction module, which focuses on capturing effective structural information from the multichannel RNA image, thereby enhancing the accuracy of RNA secondary-structure prediction. Experimental results indicate that our method achieves state-of-the-art performance compared with recent deep learning methods across several benchmark datasets and demonstrates potential in drug discovery applications.}, archive = {J_EAAI}, author = {Xiao Wang and Yongfeng Zhang and Lixiang Yang and Rong Wang}, doi = {10.1016/j.engappai.2025.112648}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112648}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Integrating multimodal biophysical features with hybrid deep learning for ribonucleic acid secondary structure prediction}, volume = {162}, year = {2025}, }
• (2025). Geometry-informed multimodal fusion network for enhancing high-density spatial transcriptomics from histology images. EAAI, 162, 112647. (www)

  @article{ , abstract = {Spatial transcriptomics (ST) is a revolutionary technology that combines spatial information with gene expression analysis, opening up new perspectives for studying the spatial distribution of genes within tissues and their regulatory mechanisms. However, it is limited by the sparsity of sequencing spots and the high cost of ST technology, which hinders its widespread application in biomedical research. An alternative and more cost-effective strategy is to leverage deep learning methods to infer high-density gene expression profiles from histological images. To this end, we developed the HisHRST (Histology-based High-Resolution Spatial Transcriptomics) method based on an offline pathological image foundation model, aiming to accurately generate high-density ST data from histological images using a geometry-guided multimodal fusion network with spatial coordinates. This method employs a multi-head attention mechanism to incorporate spatial location information, thereby enhancing feature representation. We systematically evaluated HisHRST on seven ST datasets and compared its performance with five existing methods. Experimental results demonstrate that HisHRST can accurately predict gene expression profiles for unmeasured spots, refine gene expression patterns, and effectively preserve the original spatial structure of gene expression. Furthermore, this method facilitates the identification of biologically meaningful pathways, thereby advancing the understanding of key biological processes. All code and public datasets used in this paper are available at https://github.com/wenwenmin/HisHRST and https://zenodo.org/records/15129356 .}, archive = {J_EAAI}, author = {Zhiceng Shi and Shuailin Xue and Wenwen Min}, doi = {10.1016/j.engappai.2025.112647}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112647}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Geometry-informed multimodal fusion network for enhancing high-density spatial transcriptomics from histology images}, volume = {162}, year = {2025}, }
• (2025). State-of-the-art of machine learning methods for fault detection and health monitoring of wind turbine system components: A comprehensive review. EAAI, 162, 112645. (www)

  @article{ , abstract = {As the world shifts from fossil fuels to sustainable energy sources, wind energy has become an increasingly important part of the global renewable energy mix. However, maintaining the efficient operation of wind turbines remains challenging. This review presents, discusses and extracts the major trends of advancements in artificial intelligence and their application in engineering solutions for diagnosing faults in wind turbine systems. It begins with an analysis of the benefits and limitations of different generator types used in wind power generation. Next, it identifies the most frequently failing components, including power converters, generators (with rotor), sensors, gearboxes and drivetrain. Furthermore, it provides an exhaustive overview of machine learning (ML) techniques, including advanced methodologies such as attention mechanisms and ensemble learning frameworks, for wind turbine fault detection and health monitoring. This review aims to show the current trends and future research directions of machine learning applications to improve the reliability, efficiency, and sustainability of wind turbine operations. It provides valuable insights for researchers, engineers, and decision-makers in the wind energy sector, helping them select suitable ML approaches and refine existing wind turbine operational strategies. It also enables us to measure the challenges that remain towards achieving highly efficient operation of wind turbines.}, archive = {J_EAAI}, author = {Abebe Wolie Yimam and Majid Vafaeipour and Maarten Messagie and Kinde Anlay Fante and Emiyamrew Minaye Molla and Tefera Mekonnen Azerefegn and Thierry Coosemans}, doi = {10.1016/j.engappai.2025.112645}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112645}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {State-of-the-art of machine learning methods for fault detection and health monitoring of wind turbine system components: A comprehensive review}, volume = {162}, year = {2025}, }
• (2025). Removal of cationic dyes from aqueous solutions using agricultural waste: Modeling the biosorption process with semantic immune plasma programming. EAAI, 162, 112640. (www)

  @article{ , abstract = {This study investigates the use of celery stalks, a low-cost agricultural waste, as biosorbent for removing crystal violet (CV), a toxic cationic dye, from aqueous solutions. The biosorption process was experimentally examined under varying conditions of potential of hydrogen (pH), contact time, biosorbent dose, initial dye concentration, and temperature. To model the relationship between these input parameters and biosorption performance, we employed symbolic regression using four automatic programming (AP) methods: Genetic Programming (GP), Artificial Bee Colony Programming (ABCP), Immune Plasma Programming (IPP), and a newly proposed semantic IPP (sIPP), which introduces a semantic-based crossover operator. The novelty of the study lies in combining agricultural waste-based biosorption with interpretable Artificial Intelligence (AI), enabling both accurate prediction and transparent understanding of the system. Experimental results revealed the presence of chemisorption in the spontaneous and endothermic biosorption process, with a maximum biosorption efficiency of 90% and a biosorption capacity of 6.9 mg/g. Among the symbolic models, sIPP achieved the highest test accuracy ( R Test 2 = 0 . 99 for capacity, 0.98 for efficiency), while also generating the simplest trees (approximate mean values: 91 and 99 nodes). This confirms sIPP’s strength in balancing predictive accuracy with model transparency. The proposed framework demonstrates strong potential for sustainable and explainable modeling in environmental engineering applications.}, archive = {J_EAAI}, author = {Nurşah Kütük and Sibel Arslan}, doi = {10.1016/j.engappai.2025.112640}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112640}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Removal of cationic dyes from aqueous solutions using agricultural waste: Modeling the biosorption process with semantic immune plasma programming}, volume = {162}, year = {2025}, }
• (2025). Graph-enhanced fault hierarchical perception for dissolved gas analysis in power transformer. EAAI, 162, 112639. (www)

  @article{ , abstract = {Dissolved Gas Analysis (DGA) is a widely used technique for diagnosing power transformer faults by monitoring the gas content in transformer oil to identify potential issues. However, existing research often treats dissolved gas samples as loosely connected independent entities, failing to explicitly model the relationships between them. Moreover, the similarity between faults caused by different factors can sometimes be higher than that between faults caused by the same factor, highlighting the need for hierarchical modeling of fault relationships to improve diagnostic performance. To address these challenges, we propose a graph-enhanced fault hierarchical perception model for DGA. In this work, a K -Nearest Neighbors ( K NN) graph is constructed by calculating the similarity between DGA samples in the transformer oil to capture potential associations between samples. To further enhance the learning of sample representations, we generate two augmented views by randomly perturbing node features and edges to capture information about different local structures in the graph. Additionally, we introduce a consistency constraint to ensure that the prediction results from the two augmented views remain consistent. Furthermore, we design a hierarchical perception-based ranking strategy that ranks fault similarities at a fine-grained level, effectively utilizing hierarchical fault information to optimize the model. Extensive experimental results demonstrate that the proposed model achieves high accuracy and strong generalization performance in fault-type prediction.}, archive = {J_EAAI}, author = {Yingyue Zhang and Huifang Ma and Yuwei Gao and Shengjiang Peng and Ruijia Zhang}, doi = {10.1016/j.engappai.2025.112639}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112639}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Graph-enhanced fault hierarchical perception for dissolved gas analysis in power transformer}, volume = {162}, year = {2025}, }
• (2025). Full autonomy in underwater robotics systems: A realistic prospect?. EAAI, 162, 112638. (www)

  @article{ , abstract = {Autonomous Underwater Systems (AUS) are transforming underwater exploration, environmental monitoring, and subsea operations by reducing reliance on human intervention. This review explores the advancements and challenges in achieving full operational autonomy in AUS, focusing on navigation, perception, communication, energy management, control systems, and decision-making frameworks. Current systems predominantly exhibit partial autonomy, performing pre-defined missions with limited adaptability. However, advancements in hybrid localisation techniques, multi-modal sensor fusion, and Artificial Intelligence (AI)-driven decision-making are pushing the boundaries of autonomy, enabling AUS to adapt dynamically to complex underwater environments. The integration of subsystems remains a central challenge. Navigation systems address positional drift through cooperative approaches, while perception systems enhance environmental understanding via sensor fusion. Communication technologies, although constrained by underwater limitations, are advancing through hybrid protocols that strike a balance between bandwidth and range. Energy management is evolving with innovations in battery technology, energy harvesting, and predictive resource allocation. Control systems, increasingly incorporating AI-based frameworks, translate mission commands into precise actions, bridging the gap between decision-making and physical execution. Decision-making frameworks synthesise inputs from all subsystems, enabling real-time prioritisation and adaptive behaviour. Despite significant progress, achieving full autonomy requires further innovation in subsystem integration, real-time processing, and environmental adaptability. This review emphasises the need for modular architectures and standardised protocols to ensure interoperability and scalability. The potential of fully autonomous underwater systems extends beyond technical achievements, offering profound implications for ocean exploration, resource sustainability, and environmental stewardship. The question posed – Full Autonomy in Underwater Systems: A Realistic Prospect? – is addressed with cautious optimism. While partial autonomy is well within reach, achieving full autonomy depends on sustained innovation across multiple domains. This manuscript provides a comprehensive roadmap for advancing AUS, laying the foundation for transformative advancements in one of Earth’s least understood frontiers.}, archive = {J_EAAI}, author = {Ali Rohan and Hamidreza Farhadi Tolie and Md Junayed Hasan and Somasundar Kannan}, doi = {10.1016/j.engappai.2025.112638}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112638}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Full autonomy in underwater robotics systems: A realistic prospect?}, volume = {162}, year = {2025}, }
• (2025). Design and implementation of deep learning-based framework for multi-class fault diagnosis in complex chemical process systems. EAAI, 162, 112630. (www)

  @article{ , abstract = {Fault diagnosis in modern chemical plants is increasingly challenging due to process complexity, nonlinearity, and high-risk operations, where undetected faults can cause severe safety and economic consequences. Conventional machine learning (ML) models suffer from reliance on handcrafted features, poor generalization in high-dimensional spaces, and limited labeled data, resulting in reduced diagnostic performance. To overcome these challenges, we propose a scalable deep learning (DL) framework for multi-class fault diagnosis in chemical processes. The framework employs convolutional neural networks (CNN), autoencoders (AE), and long short-term memory (LSTM) networks to automatically extract spatial and temporal features from multivariate process data. Validated on the Tennessee Eastman process (TEP) benchmark, CNN achieved the highest diagnostic performance, with 88 % accuracy, 91 % precision, and 89 % F1-score. AE also performed strongly, with 85 % accuracy and 82 % F1-score, while LSTM achieved 71 % accuracy and 75 % F1-score, all outperforming conventional machine learning models, which scored between 48 % and 52 % accuracy. Superior AUC scores (micro-average: 1.00; macro-average: 0.99) confirm the framework's robustness, including in detecting overlapping and rare faults. The proposed two-phase approach, offline training and real-time monitoring, offers a practical solution for improving fault diagnosis accuracy, adaptability, and early warning capabilities in industrial processes.}, archive = {J_EAAI}, author = {Remigius Nnadozie Ewuzie and Shivaneswar Gunasekaran and Zainal Ahmad and Norazwan Md Nor}, doi = {10.1016/j.engappai.2025.112630}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112630}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Design and implementation of deep learning-based framework for multi-class fault diagnosis in complex chemical process systems}, volume = {162}, year = {2025}, }
• (2025). A neural network approach on forecasting the engine emission using biodiesel fuel from fish oil based on injection pressure and engine operating parameters. EAAI, 162, 112628. (www)

  @article{ , abstract = {This study investigates the influence of injection pressure on the performance and emission characteristics of a diesel engine fueled with biodiesel derived from basa fish oil blended with conventional diesel (B10, B20, and B30). Experiments were performed at injection pressures ranging from 400 to 800 bar, with 600 bar as the baseline, on an engine operating at 1400 revolutions per minute (rpm). The results indicated that reducing the injection pressure to 500 bar increased emissions of carbon monoxide (CO), hydrocarbons (HC), and smoke, while nitrogen oxides (NOx) emissions decreased. At 500 bar compared to 600 bar, CO, HC, and smoke rose by up to 23.4 %, 10.0 %, and 32.0 %, respectively, whereas NOx decreased by up to 9.1 %. Conversely, at 700 bar, CO, HC, and smoke emissions decreased by up to 20.3 %, 6.7 %, and 28.4 %, while NOx increased by up to 12.7 %. Furthermore, an exhaust emission prediction model was developed using an Artificial Neural Network (ANN) based on injection pressure and engine operating parameters. The ANN model achieved high predictive accuracy, with the coefficient of determination (R 2 ) exceeding 0.995 and root mean square error (RMSE) values as low as 0.49975 % for smoke. These findings demonstrate that injection pressure significantly affects emission trade-offs in diesel engines operating with biodiesel blends. In addition, the ANN model proves to be a reliable tool for predicting and controlling engine emissions, supporting the potential of biodiesel as a sustainable alternative fuel.}, archive = {J_EAAI}, author = {Nguyen Van Tuan and Nguyen Xuan Khoa and Nguyen Tuan Nghia}, doi = {10.1016/j.engappai.2025.112628}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112628}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A neural network approach on forecasting the engine emission using biodiesel fuel from fish oil based on injection pressure and engine operating parameters}, volume = {162}, year = {2025}, }
• (2025). A fuzzy decision-making method for unloading command at crushing stations based on deep learning and dynamic coal flow features slicing. EAAI, 162, 112617. (www)

  @article{ , abstract = {The contemporary world hosts numerous open-pit coal mines, as mining efficiency increases, the unloading process of the crushing station still relies on manual command. Adversely affected by the harsh production environment and the high-intensity fatigue of operators, incorrect unloading commands can lead to issues, such as blockages at the receiver bin discharge outlet and coal swelling in the crushing chamber, resulting in production accidents. To address this issue, this study analyzes the dynamic characteristics of the coal flow in the receiver bin, in addition, monitors the bin status in real time through the camera, combining a lightweight selective kernel network (Li-SKNet) with the average classification confidence from a small segment of consecutive video frames to assess the suitability of unloading conditions. By introducing a convolutional kernel attention mechanism, the model achieved a classification accuracy of 99.8 %, subsequently, in order to adapt to the changes in the working conditions of the crushing station, a segmented fuzzy function is proposed to further optimize the model inference results. Finally, an intelligent unloading command system for the crushing station is established, which has been put into production and operated stably for six months. In comparison to traditional manual commanding methods for unloading, the system has approximately demonstrated a 15 % increase in production efficiency.}, archive = {J_EAAI}, author = {Tongyu Cui and Yongtai Pan and Yankun Bi and Zhen Liu and Jiacheng Huang and Bingjia Liu}, doi = {10.1016/j.engappai.2025.112617}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112617}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A fuzzy decision-making method for unloading command at crushing stations based on deep learning and dynamic coal flow features slicing}, volume = {162}, year = {2025}, }
• (2025). Predicting the cyclic behavior and strength-weakening effect of saturated clays using artificial neural network. EAAI, 162, 112616. (www)

  @article{ , abstract = {The cyclic behavior of soft cohesive soils under shear loads is characterized by progressively increasing strain and by the growth of the pore pressure that can lead to an effective stress reduction and, eventually, to the sudden failure of the soil, both risks with evident engineering safety implications. Although this problem has received much attention in research, most present approaches can only predict some parameters of the clay performance separately, commonly leading to highly complex approaches requiring extensive training and expertise. The present research uses an Artificial Neural Network (ANN) and machine learning to predict the cyclic behavior of the soft clays in the investigated site, considering for the first time all the relevant parameters that characterize the problem. The proposed ANN includes 9 inputs, two hidden layers with 10 neurons each, and five outputs. Nine inputs include the vertical effective consolidation pressure, parameters from the monotonic shear test, and defined input variables from the cyclic simple shear test. As outputs, the net considers five different results, including the various parameters of the shear strain response for each cycle, the maximum number of cycles, and the pore pressure increase. The resulting ANN shows predictions with high accuracy, with R = 0.995, and individual errors below 10 % in most cases. No prior training or experience is required to use the ANN, and it can be confidently used as an alternative to other analytical and numerical approaches for analyzing clay cyclic behavior, as long as the input values fall within the defined ranges.}, archive = {J_EAAI}, author = {M.A. Millán and R. Galindo and A. Viana da Fonseca and H. Patiño}, doi = {10.1016/j.engappai.2025.112616}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112616}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Predicting the cyclic behavior and strength-weakening effect of saturated clays using artificial neural network}, volume = {162}, year = {2025}, }
• (2025). Synergistic co-evolution with neural networks for evolutionary optimization. EAAI, 162, 112614. (www)

  @article{ , abstract = {Evolutionary computation (EC) algorithms, which simulate natural selection and genetic mechanisms, are widely applied to solve complex optimization and search problems. During the evolutionary process, EC algorithms inherently generate a wealth of evolutionary data. However, conventional EC algorithms often fail to effectively utilize this data, limiting their optimization effectiveness and overall performance. Neural networks, known for their powerful learning capabilities, excel in extracting features and recognizing patterns within large datasets. They can automatically identify complex relationships through their hierarchical architectures. Inspired by this, we propose a neural-synergized co-evolutionary optimization (NSCO) framework that integrates neural networks to learn from successfully evolved individuals during the EC process. This approach extracts valuable evolutionary knowledge to guide algorithms toward improved solutions and higher-quality data. Additionally, the enhanced data enables neural networks to derive richer evolutionary insights, creating a positive feedback loop that consistently improves performance. Notably, this framework operates without the need for additional expert knowledge, relying solely on the data generated by the algorithms themselves. To validate its effectiveness, we integrate this framework with 10 distinct EC algorithms and evaluate its performance using the CEC2014 benchmark suite. Results demonstrate significant enhancements in the algorithms’ performance.}, archive = {J_EAAI}, author = {Kun Bian and Juntao Zhang and Hong Han and You Zhou and Yifei Sun and Shi Cheng and Jun Zhou}, doi = {10.1016/j.engappai.2025.112614}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112614}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Synergistic co-evolution with neural networks for evolutionary optimization}, volume = {162}, year = {2025}, }
• (2025). Evaluation of environmental emergency treatment technologies using the interval pythagorean neutrosophic set. EAAI, 162, 112611. (www)

  @article{ , abstract = {At present, the problem of environmental pollution is severe both domestically and internationally, hindering social development. Therefore, strengthening environmental protection and improving the ability to manage environmental emergencies are urgently needed. Based on this situation, this paper proposes an evaluation model for environmental emergency treatment techniques (EETTs) using the interval Pythagorean neutrosophic (IPN) set for the first time. First, the IPN is introduced, and the operating rules and fractional/exact functions of IPN numbers are defined. Second, to apply IPN numbers to multi-attribute decision-making, two types of aggregation operators (IPN weighted arithmetic average (IPNWAA) operator and IPN weighted geometric average (IPNWGA) operator) are proposed. Finally, the evaluation index system and model for EETTs are constructed, and the rationality of the model is verified through an example involving computer data processing. According to the evaluation results, enterprises can identify their strengths and weaknesses in coping with environmental emergency treatment and take corresponding corrective measures in time to improve their emergency treatment capabilities. Therefore, this study has important scientific value and practical significance.}, archive = {J_EAAI}, author = {Changxing Fan}, doi = {10.1016/j.engappai.2025.112611}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112611}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Evaluation of environmental emergency treatment technologies using the interval pythagorean neutrosophic set}, volume = {162}, year = {2025}, }
• (2025). A dynamic consensus-based decentralized method for multiple unmanned aerial vehicles cooperative target allocation. EAAI, 162, 112609. (www)

  @article{ , abstract = {This paper investigates multiple unmanned aerial vehicles dynamic target allocation problem with capability-requirement matching constraints. The objective is to achieve conflict-free allocation through distributed collaboration under a dynamic environment. To address this problem, we propose a decentralized method named Dynamic Consensus-Based Group Algorithm (DCBGA). This algorithm utilizes a rule-based method known as State-Event-Condition-Action (SECA) as the decision-making framework. In addition, this algorithm consists of iterations between two phases: target selection and conflict resolution. In the target selection phase, each unmanned aerial vehicle (UAV) selects targets in a market-based method and follows the principle of “capability-requirement matching”. In the conflict resolution phase, a consensus-based mechanism is designed so that each UAV communicates with its neighbors to generate a conflict-free allocation result. In the simulation section, this paper performs convergence analysis, sensitivity analysis, and optimization comparisons. The simulation results confirm both the convergence and robustness of the DCBGA. Furthermore, the comparison results demonstrate that the proposed algorithm exhibits better optimization performance than other advanced algorithms.}, archive = {J_EAAI}, author = {Han Wang and Xiaolong Liang and Jiaqiang Zhang and Aiwu Yang}, doi = {10.1016/j.engappai.2025.112609}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112609}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A dynamic consensus-based decentralized method for multiple unmanned aerial vehicles cooperative target allocation}, volume = {162}, year = {2025}, }
• (2025). Biological visual-cognition-inspired deep network for short-term significant wave height prediction. EAAI, 162, 112607. (www)

  @article{ , abstract = {Accurate prediction of Significant Wave Height (SWH) is essential to optimize wave energy conversion efficiency. However, ocean waves' stochastic and nonlinear characteristics make traditional models face the challenges of insufficient accuracy and limited interpretability. In this paper, we propose a deep network based on a biological visual cognitive mechanism, which significantly improves the short-term SWH prediction performance inspired by the functional principles of the human visual system's hierarchical perception, hemispheric collaboration and attentional decision-making mechanism. The model consists of three components: (1) a visual-spatial pyramid component, which gradually extracts local to global features through a multi-scale convolutional kernel; (2) a brain analysis component, which combines gated recurrent unit (GRU) and convolutional neural network (CNN) to capture spatiotemporal dependencies and enhance the stability through residual connectivity, respectively; and (3) an attention-driven prediction component, which dynamically filters the key features to improve the prediction accuracy. Experiments based on two real-world buoy sites show that the proposed method reduces the mean absolute percentage error (MAPE) to 5.56 % and 6.30 %, respectively. The energy capture error (ΔP) is reduced to 10.93 % and 13.02 %, respectively, and remains robust under extreme sea states. This work can effectively improve the prediction accuracy and operational reliability of industrial ocean energy systems under complex ocean conditions.}, archive = {J_EAAI}, author = {Liao Fang and Weimin Wu and Feifei Cao and Zhenguan Cao and Guodong Fan and Lin Cui and Frede Blaabjerg}, doi = {10.1016/j.engappai.2025.112607}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112607}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Biological visual-cognition-inspired deep network for short-term significant wave height prediction}, volume = {162}, year = {2025}, }
• (2025). Event-based video interpolation via complementary motion information. EAAI, 162, 112606. (www)

  @article{ , abstract = {Video frame interpolation, the task of synthesizing intermediate frames to increase temporal resolution, often struggles with complex scenarios when constrained by the assumption of linear motion The advent of event cameras has led to significant progress in addressing this issue. Event cameras, with microsecond-level temporal resolution, bridge the gap between frames by providing accurate motion cues. However, current event-based video frame interpolation methods often overlook that event data primarily offers high-confidence features at scene edges during multi-modal feature fusion, which may limit the contribution of event signals to optical flow estimation. To address this, we propose a novel end-to-end learning framework that explicitly leverages the complementary characteristics of event signals and frames. Our method synergistically fuses dense contextual information from frames with sparse but precise edge motion from events via a proposed Edge Guided Attention (EGA) module. The EGA employs a coarse-to-fine strategy, where event-based optical flow directly refines the frame-based motion estimation at each level of a pyramidal architecture. Additionally, we introduce an event-based visibility map, co-learned within our event-processing network, to adaptively mitigate occlusions during the warping process. Extensive experiments conducted on a diverse suite of six benchmarks, including four synthetic and two real-world datasets validate the effectiveness of this novel approach. A dedicated discussion of the method’s trade-offs and potential limitations is presented in the Limitations section.}, archive = {J_EAAI}, author = {Yuhan Liu and Linghui Fu and Hao Chen and Zhen Yang and Youfu Li and Yongjian Deng}, doi = {10.1016/j.engappai.2025.112606}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112606}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Event-based video interpolation via complementary motion information}, volume = {162}, year = {2025}, }
• (2025). Real-time dynamic coordinated optimization control with near-global optimal learning for connected plug-in hybrid electric vehicles. EAAI, 162, 112602. (www)

  @article{ , abstract = {With the application of connected technologies, such as vehicle-to-vehicle and vehicle-to-cloud communication in connected vehicles to obtain various traffic information, it is crucial to balance the optimality and computational burden of the energy management strategy for further improving the fuel economy of the connected plug-in hybrid electric vehicle. Another key factor affecting the improvement of fuel economy and control performance in connected plug-in hybrid electric vehicles is the fluctuation in driving torque resulting from the different response characteristics of the engine and motor during mode switching of the powertrain for the power distribution of the energy management strategy. To address these challenges, this paper proposes a novel real-time dynamic coordinated optimization control scheme that incorporates energy management at the upper layer and adaptive coordination at the lower layer for the connected plug-in hybrid electric vehicle in a vehicle-following scenario. Based on the offline optimal control rules extracted by the extreme learning machine, which possesses good generalization capabilities, the upper-layer guided model predictive control for energy management is implemented by applying the particle swarm optimization algorithm within variable horizons across different road sections. The bottom-layer adaptive fixed-time control scheme, equipped with a coordinated mechanism, is designed to address transient response deviations in the upper-layer results. The effectiveness and advantages of the proposed hierarchical scheme are validated through both the co-simulation platform and a hardware-in-loop test.}, archive = {J_EAAI}, author = {Jiaqi Xue and Chao Yang and Jiayi Fang and Xiao Zhang and Muyao Wang}, doi = {10.1016/j.engappai.2025.112602}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112602}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Real-time dynamic coordinated optimization control with near-global optimal learning for connected plug-in hybrid electric vehicles}, volume = {162}, year = {2025}, }
• (2025). A multichannel framework for multitask learning fusion in modulation recognition tasks. EAAI, 162, 112601. (www)

  @article{ , abstract = {Automatic modulation recognition is an essential process linking signal detection to signal demodulation, and it is a promising technology for improving spectrum usage efficiency in cognitive radio. With the advent of fifth generation mobile communication technology, wireless communication systems have had massive data throughput. As a result, integrating artificial intelligence techniques with modulation recognition has emerged as a key focus area in the communication sector. To enhance the precision of signal modulation recognition, this study proposes a multitask learning fusion multichannel network modulation recognition framework—Multitask Residual Convolution Long Short Memory-Transformer Deep Neural Network. The primary task network is a two stream network composed of Convolutional Neural Network module, Long Short-Term Memory module and Transformer-Encoder module, allowing for the simultaneous extraction of both time and frequency features from the signal. The auxiliary task network is composed of residual convolution module paired with Transformer-Encoder module, designed for extracting the power spectral density characteristics of straightforward signals. Ultimately, the features derived from both the primary task network and the auxiliary task network are combined. The auxiliary network enhances the primary task network’s ability to characterize features, thereby boosting the neural network’s overall versatility and precision. The experimental findings indicate that the proposed model achieved peak recognition accuracies of 99.73%, 93.9%, and 94.1% on the three datasets, respectively. Moreover, the recognition accuracy of the proposed model is better than the baseline model in the low Signal-to-Noise Ratio environment (-18 decibel ∼ 0 decibel).}, archive = {J_EAAI}, author = {Wenshi Xiao and Zhongqiang Luo and Mengxuan Lan}, doi = {10.1016/j.engappai.2025.112601}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112601}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multichannel framework for multitask learning fusion in modulation recognition tasks}, volume = {162}, year = {2025}, }
• (2025). Energy-efficiency prediction of fluid machinery based on knowledge-aided ensemble learning. EAAI, 162, 112600. (www)

  @article{ , abstract = {Fluid machinery is a type of high-energy-consumption equipment. Accurate prediction of energy efficiency enables the optimization of energy management and enhances equipment performance. However, factors such as equipment aging, sensor anomalies, and complex interferences pose challenges to accurately predicting energy efficiency solely through knowledge-based models or machine learning approaches. To address this issue, a fluid machinery energy efficiency prediction method based on knowledge-aided ensemble learning is proposed by integrating mechanistic knowledge and machine learning. During data repair, the physical characteristics of fluid machinery are used to identify condition anomalies, and a sliding window-based Local Outlier Factor (WLOF) algorithm is designed to identify other types of anomalies. Different types of anomalies are repaired using mechanistic knowledge and eXtreme Gradient Boosting (XGBoost), respectively. The preprocessed data is subsequently employed to train models for energy efficiency prediction. This method has been applied to predict the energy efficiency of fans in a large-scale steel plant. Compared with various energy efficiency prediction models, the application results show that the prediction based on Knowledge-WLOF-XGBoost exhibits higher accuracy and robustness than single physical models or machine learning models.}, archive = {J_EAAI}, author = {Weijian Kong and Rong Zhuo and Linfu Zheng}, doi = {10.1016/j.engappai.2025.112600}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112600}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Energy-efficiency prediction of fluid machinery based on knowledge-aided ensemble learning}, volume = {162}, year = {2025}, }
• (2025). Pretrained transformers for multimodal fake news detection: Explainability using SHapley additive exPlanations for contributions from text, image, and image captions. EAAI, 162, 112590. (www)

  @article{ , abstract = {The wide use of Online Social Networks (OSNs) for news consumption necessitates the identification and flagging of fake news to reduce its negative impacts. Existing research mostly emphasizes deep learning techniques that examine textual features or combinations of text and images for the detection of fake news. In this study, we propose two enhanced models for fake news detection. First, we present a multimodal framework that leverages Vision Transformer embeddings for image representation instead of traditional Convolutional Neural Network (CNN) features, which achieves significant improvements over baseline approaches. Second, we implemented an image caption-augmented multimodal model that integrates image captions as an additional modality alongside text and images. Caption embeddings act as semantic bridges that strengthen cross-modal alignment and contextual understanding. The experiments conducted on the FakeNewsNet and Fakeddit datasets show that the model incorporating image captions consistently surpasses the baseline BERT-VIT framework. It achieves improvements of up to +7.9% in accuracy, +11.5% in precision, and +5.3% in the F1-score, while still maintaining competitive levels of recall and Area Under the Curve (AUC). Statistical significance testing verifies the reliability of these enhancements. Additionally, a SHapley Additive exPlanations (SHAP)-based analysis highlights the contribution of each modality, which enhances the interpretability of the proposed framework.}, archive = {J_EAAI}, author = {Athira A.B. and S.D. Madhu Kumar and Anu Mary Chacko}, doi = {10.1016/j.engappai.2025.112590}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112590}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Pretrained transformers for multimodal fake news detection: Explainability using SHapley additive exPlanations for contributions from text, image, and image captions}, volume = {162}, year = {2025}, }
• (2025). Assessing the morphological quality of zizania latifolia shoot based on instance segmentation. EAAI, 162, 112589. (www)

  @article{ , abstract = {The breeding of Zizania latifolia primarily relies on the cross-regional introduction of varieties and manual selection, which involves the assessment of morphological quality. Traditional methods are time-consuming and labor-intensive. This paper proposed a morphological assessment method based on instance segmentation. Assessment indicators, including enlargement degree, color, straightness, and slenderness, were defined according to national grading standards and expert knowledge to provide a comprehensive assessment framework. For the segmentation task, the irregular contours, similar internal regions, and indistinct boundaries pose significant challenges. Additionally, classification heavily relies on the spatial relationships. To solve these problems, this study proposed a sub-stem segmentation algorithm, which based on You Only Look Once Version 8 (YOLOv8). The Linear Deformable Convolution (LDConv) module was employed to recognize irregular sub-stems through learnable offsets. Next, the Efficient Multi-Scale Attention (EMA) mechanism was added to refine image features by emphasizing the interactions between channels and spatial dimensions. We also employed the Wise Intersection over Union Version 3 (WIoU v3) loss function, which evaluated the anchor box quality based on outlier degree. The proposed model outperformed YOLOv8 and other state-of-the-art models on the Zizania latifolia dataset, achieving Precision, Recall, and mean Average Precision (mAP) values of 95.3 %, 93.9 %, and 97.7 %. Additionally, our model achieved 35.7 % mAP on the Microsoft Common Objects in COntext (MS COCO) dataset. Based on the segmentation results, we extracted phenotypic parameters, such as height, width, and two-dimensional area. We determined the top-10 individuals with superior morphological quality among 65 varieties, providing robust support for Zizania latifolia breeding.}, archive = {J_EAAI}, author = {Chenmin Yang and Shanyong Wang and Tingting Lou and Ruiqi Song and Huanliang Xu and Zhaoyu Zhai}, doi = {10.1016/j.engappai.2025.112589}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112589}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Assessing the morphological quality of zizania latifolia shoot based on instance segmentation}, volume = {162}, year = {2025}, }
• (2025). A combined adaptive gaussian short-term fourier transform and mamba framework for stock price prediction. EAAI, 162, 112588. (www)

  @article{ , abstract = {In financial investment, predicting stock prices precisely is challenging due to the inherent volatility and non-stationarity of market data. To address this, we propose a pioneering hybrid approach, introducing the Adaptive Gaussian Short-Time Fourier Transform (AG-STFT) in Mamba for the first time—it dynamically adjusts the window size based on frequency content to enhance the representation of non-stationary modes. We further present AGSMNet, a novel hybrid architecture fusing AG-STFT with the Mamba framework, which addresses the dual-domain modeling challenge (time + frequency) through an end-to-end deep learning solution leveraging both signal processing and structured state-space modeling techniques, enabling robust sequence modeling with enhanced temporal and spectral sensitivity. The AG-STFT dissects and represents temporal and spectral features of stock price movements, capturing dynamics often overlooked by traditional models, while Mamba — known for discerning complex patterns — enhances predictive capabilities. The model employs a Multiple Feature Extraction (MFE) module for shallow features and a Residual State-Space Group (RSSG) module for deep features, fed into a predictor module to predict next-day prices. Extensive experiments across ten major stock and index datasets demonstrate that AGSMNet consistently achieves the lowest MAE, MSE, and RMSE, and the highest R 2 values, outperforming leading models such as TimesNet, Autoformer, and MambaStock. These results confirm AGSMNet’s superior generalization, robustness across volatile markets, and practical effectiveness. It is applicable to non-stationary datasets, outperforms traditional, machine learning, and deep learning-based methods, and shows potential as a transformative tool for financial analysts and investors, with adaptability to other financial modeling areas promising advancements in predictive analytics.}, archive = {J_EAAI}, author = {Yuling Huang and Zhiyuan Pei and Jin Yan and Chujin Zhou and Xiaoping Lu}, doi = {10.1016/j.engappai.2025.112588}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112588}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A combined adaptive gaussian short-term fourier transform and mamba framework for stock price prediction}, volume = {162}, year = {2025}, }
• (2025). Driver risk classification for transportation safety: A machine learning approach using psychological, physiological, and demographic factors with driving simulator. EAAI, 162, 112585. (www)

  @article{ , abstract = {Driving behavior is a critical factor in ensuring road safety, particularly in commercial transportation sectors where hiring safe and reliable drivers is a priority. This study presents a machine learning-based framework for predicting driver behavior using multimodal assessments, including psychological, physiological, and demographic factors. We utilized a driving simulator equipped with biometric sensors to capture the physiological data of the driver, including heart rate, eye blink rate, pupil diameter, and point of gaze (POG), during driving sessions. The psychological attributes are collected through a self-report questionnaire. The questionnaire is structured to obtain information about nine psychological characteristics of the driver, including instrumental attitude, social anxiety, sensation seeking, premeditation, urgency, selfishness, aggressive mode, life satisfaction, and conscientiousness. In addition, some demographic attributes, such as age and gender, are also adopted to study their effect on driving behavior. Experiments were conducted on 80 participants, each driving for 10 min. Various machine learning models, along with a feature selection strategy, were used to find the relationship between the driver's modalities and his driving behavior. Results demonstrate that the k-nearest neighbors (KNN) model achieved the best performance, yielding an accuracy of 93.75 % and a False Negative Rate (FNR) of 0. Feature importance analysis revealed that gaze distraction, sensation seeking, conscientiousness, and gender are the best predictors of driving behavior. The findings suggest that our model can serve as a valuable decision-support tool for taxi companies and transportation agencies aiming to enhance driver selection processes by identifying drivers with lower accident risks.}, archive = {J_EAAI}, author = {Malek Masmoudi and Yasmin Shakrouf and Omar Hassan Omar and Amir Shikhli and Fatima Abdalla and Wadad Alketbi and Imad Alsyouf and Ali Cheaitou and Anwar Jarndal and Ali I. Siam}, doi = {10.1016/j.engappai.2025.112585}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112585}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Driver risk classification for transportation safety: A machine learning approach using psychological, physiological, and demographic factors with driving simulator}, volume = {162}, year = {2025}, }
• (2025). A credibility and consistency-oriented stochastic aggregation framework for heterogeneous multi-attribute large-scale group decision making with several attribute sets. EAAI, 162, 112584. (www)

  @article{ , abstract = {In large-scale group decision making, experts usually provide their individual preferences on attribute values, as well as attribute sets. The coexistence of large-scale heterogeneous information poses a significant challenge to the credibility and consistency on decision making. To this issue, the paper proposes a credibility and consistency-oriented stochastic aggregation framework, including four primarily research points. Firstly, a simplified transformation method is developed to convert heterogeneous attribute values into individual attribute superiority-probability-based pairwise comparison matrix (IA-SPMs), which saves transformation cost as well as provides abundant references for credibility analysis. Secondly, the deviation-based credibility measures and the credibility-based weighting methods are proposed. Thirdly, a cluster-based aggregation operator is introduced by considering group consistent preferences on attributes selection to get the collective SPM (C-SPM). Fourthly, the ranking probability matrix (RPM) and the possibility ranking result are calculated based on the C-SPM. Using experiment and application analyses we illustrate that the proposed methods can enhance the credibility of decision outcomes, as well as the stability of the results. This research can provide technical support for effective fusion of heterogeneous information from multiple sources in the artificial intelligence (AI) era, and has broad application potential in large-scale democratic decision making and pre-evaluation engineering projects.}, archive = {J_EAAI}, author = {Weiwei Li and Pingtao Yi and Danning Zhang}, doi = {10.1016/j.engappai.2025.112584}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112584}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A credibility and consistency-oriented stochastic aggregation framework for heterogeneous multi-attribute large-scale group decision making with several attribute sets}, volume = {162}, year = {2025}, }
• (2025). Probabilistic machine learning with a cascaded framework for robust failure mode classification and capacity estimation of rectangular concrete-filled steel tube columns under axial compression. EAAI, 162, 112581. (www)

  @article{ , abstract = {This study develops an advanced probabilistic machine learning (ML) framework to predict failure modes and axial load-bearing capacity of rectangular concrete-filled steel tube (CFST) columns, addressing critical challenges of class imbalance and prediction uncertainty. Leveraging a comprehensive dataset of 597 experimental samples, a rigorous feature selection procedure is conducted to optimize the ML applications. Three ML models are developed and evaluated: a deterministic approach (Random Forest) and two probabilistic models (Gaussian Process and Natural Gradient Boosting, NGBoost). To address the issue of class imbalance in failure mode classification, a hybrid resampling strategy combining the Synthetic Minority Over-sampling Technique and Tomek Links method (SMOTE-Tomek) is implemented. Furthermore, A novel classification-regression cascaded framework is proposed, where failure modes are first classified, followed by category-specific regression for capacity prediction. Results demonstrate that SMOTE-Tomek significantly improves minority failure mode classification, increasing the average F1-scores by 8.1 % (for flexural failure) and 118.2 % (for combined failure). The cascaded framework outperforms direct regression in estimating capacity, achieving a 56 % higher coefficient of determination (R 2 ) and 15–30 % lower error metrics on average. NGBoost excels in both probabilistic failure mode prediction and uncertainty-aware capacity estimation, enabling reliability-based design and risk-informed decision-making for CFST structures in engineering practice.}, archive = {J_EAAI}, author = {Xueqi Zhong and Dade Lai and Qiyao Yu and Feiyu Liao}, doi = {10.1016/j.engappai.2025.112581}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112581}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Probabilistic machine learning with a cascaded framework for robust failure mode classification and capacity estimation of rectangular concrete-filled steel tube columns under axial compression}, volume = {162}, year = {2025}, }
• (2025). Enhancing precision tomato harvesting with feature sharing and intersection screening. EAAI, 162, 112580. (www)

  @article{ , abstract = {Tomato is a significant economic crop cultivated extensively worldwide. Rapid and accurate localization of picking points is fundamental to automating the tomato harvesting process. To address this, we develop a lightweight dual-task network built upon the YOLOv10 (You Only Look Once) architecture. The backbone integrates the ADown module for efficient downsampling, and utilizes a Dysample-based cross-scale feature fusion module (CCFM) in the detection neck, while the segmentation neck employs Simulated Annealing Adaptive Concatenation (SAAC) for feature fusion. Furthermore, the conditionally restricted Hungarian algorithm is employed to match ripe tomatoes with their corresponding corollas. Two localization techniques are proposed: local skeleton line fitting and boundary intersection screening. In local skeleton line fitting, the corolla detection box is expanded, and the stem mask within this region is used to extract the skeleton line and find the intersection point. Boundary intersection screening determines whether to directly compute the midpoint of intersecting lines or use skeleton line fitting, based on the number of intersections between the stem mask and the corolla box. The detection Precision and Recall of the proposed model are 0.970 and 0.983, respectively, while the IoU (Intersection over Union) and MIoU (Mean Intersection over Union) for segmentation are 0.962 and 0.980. The localization precision for picking points is 0.922, with a recognition rate of 0.988. Achieving an average processing time of only 55.89 ms per image, the proposed artificial intelligence application exhibits excellent efficiency and accuracy in real-world conditions, offering a promising foundation for further development of automated tomato harvesting systems.}, archive = {J_EAAI}, author = {Xueke An and Bin Dai and Zhaolei Yang and Hassan H.A. Mostafa and Wang Fangyan and Yuliang Yun}, doi = {10.1016/j.engappai.2025.112580}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112580}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing precision tomato harvesting with feature sharing and intersection screening}, volume = {162}, year = {2025}, }
• (2025). Predicting the compression index of expansive soils with hybrid machine learning approaches. EAAI, 162, 112579. (www)

  @article{ , abstract = {Expansive soils pose significant challenges for geo-infrastructure design, construction, and maintenance because of their moisture-induced volume changes. Despite extensive research on swelling behavior, the compression index ( C c ), an important indicator of soil compressibility has received relatively limited attention. C c is conventionally obtained from time-consuming and costly consolidation tests, while empirical equations derived from general clays may not provide reliable estimates for expansive soils. To address this gap, seven machine learning models based on five algorithms were developed to estimate the C c of expansive soils using soil properties readily obtained from conventional laboratory tests. A comprehensive dataset comprising 238 expansive soil samples, compiled from 60 years of published literature across various regions of the world, is employed to train and validate the proposed models. Among them, Model 2 based on Newton-Raphson-Based Optimizer optimized Extreme Gradient Boosting (NRBO-XGBoost) achieved the best performance, with a coefficient of determination ( R 2 ) of 0.903 and a root mean square error (RMSE) of 0.029 on the test set. Sensitivity analysis showed that the plasticity index was the most influential factor (31.1 %), followed by liquid limit (29.4 %), initial void ratio (25.3 %), and dry density (14.1 %), highlighting its primary influence on soil compressibility. Additionally, a simplified equation derived from the Multilayer Perceptron (MLP) designated as Model 4 is validated through three case studies. The settlement predictions deviated by 5–14 % from field measurements, offering a practical tool without the need for machine learning techniques. The findings provide useful guidance for developing rational design strategies for geo-infrastructures affected by expansive soils.}, archive = {J_EAAI}, author = {Aolin Zhang and Sai K. Vanapalli}, doi = {10.1016/j.engappai.2025.112579}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112579}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Predicting the compression index of expansive soils with hybrid machine learning approaches}, volume = {162}, year = {2025}, }
• (2025). Model-based pre-compensator design for piezoelectric actuators based on physics guided neural network and physics-precision balanced training. EAAI, 162, 112576. (www)

  @article{ , abstract = {Designing model-based pre-compensators is crucial for achieving high-precision motion control in piezoelectric actuators (PEAs), which exhibit complex nonlinear behaviors. While data-based modeling methods offer a straightforward approach, they face challenges with poor extrapolation and high complexity. Incorporating the known physics of PEAs, which is more consistent with physical principles and computationally efficient, can help overcome these limitations. Based on the above knowledge, this paper embeds physical knowledge within a neural network, forming a novel Physics Guided Neural Network (PGNN) structure as a model-based PEA pre-compensator to achieve high motion precision. Training the PGNN is challenging due to potential competition between the physics component and the neural network. The flexible nature of the neural network can easily overshadow the physical information, leading to overfitting and rendering the PGNN model ineffective. To address this, a Physics-Precision Balanced Training (PPBT) method is proposed. In the PPBT method, the physical correctness and model precision of the PGNN are mathematically defined, and these two components are balanced through a nonlinear function within the training algorithm. Experimental results show that the proposed pre-compensator based on PGNN and PPBT outperforms physics-based approaches in precision and offers greater robustness than purely data-based methods. The peak-to-peak displacement error is reduced to less than 35 nm in open-loop control. Measured by Mean Absolute Error (MAE), this method reduces displacement errors by 89 % compared to no compensation, by 77 % compared to purely neural network-based compensation, and by 73 % compared to rate-dependent Prandtl-Ishlinskii operator-based compensation.}, archive = {J_EAAI}, author = {Qin Li and Zhiwei Ruan and Chenyang Ding}, doi = {10.1016/j.engappai.2025.112576}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112576}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Model-based pre-compensator design for piezoelectric actuators based on physics guided neural network and physics-precision balanced training}, volume = {162}, year = {2025}, }
• (2025). End-to-end burst signal demodulation via adaptive masked deep learning framework. EAAI, 162, 112569. (www)

  @article{ , abstract = {The bit error rate (BER) directly determines the quality of wireless communication transmission. Traditional demodulators are limited in operating on burst signals and exhibit poor BER performance in low signal-to-noise ratio (SNR) conditions. For real-world burst signals, symbol-by-symbol approaches fail to capture inter-symbol dependencies, and existing end-to-end frameworks cannot handle the variable output lengths required for burst signals. To address this issue, we propose an end-to-end demodulation framework based on deep learning (DL), in which detection, recognition, channel compensation, and demodulation stages were trained as a unified system, enabling the entire signal burst to be demodulated in a single operation during inference. The framework's generalization and robustness are enhanced by a proposed masking mechanism and a denoising autoencoder (DAE), respectively. The former dynamically adjusts the output bitstream length while preventing gradient flow from redundant components, and the latter compensates for channel fading effects. We further introduce a dedicated end-to-end training strategy to optimize the adaptation between these modules. Experimental results on real-world Frequency Shift Keying (FSK), Minimum Shift Keying (MSK), Phase-Shift Keying (PSK), and Quadrature Amplitude Modulation (QAM) signals demonstrate that the proposed framework achieves superior demodulation accuracy for long-sequence burst signals. Compared to existing methods, the proposed framework enables parallel demodulation, and dynamically adapts the output bit stream in terms of varying message types and lengths.}, archive = {J_EAAI}, author = {Mingdi Li and Wenzhe Fan and Yanbin Li and Chunlei Xie and Yanan Duan}, doi = {10.1016/j.engappai.2025.112569}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112569}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {End-to-end burst signal demodulation via adaptive masked deep learning framework}, volume = {162}, year = {2025}, }
• (2025). Dual-head detector with point-driven transformer and semantic-spatial gating for liquid crystal display defects. EAAI, 162, 112565. (www)

  @article{ , abstract = {The unique structure of liquid crystal displays (LCDs) leads to defects with specific characteristics, posing a significant challenge to defect detection. These defects mainly include global slender line defects, numerous small defects, and other defects exhibiting significant variations in scale and visual complexity. To simultaneously detect slender defects and numerous small defects, we propose a novel dual-head detector (DHDet) that combines the strengths of transformer-based and convolutional neural network (CNN)-based approaches. The transformer-based line head leverages its global context modeling capability and exploits shape priors by representing line defects using discrete points. Additionally, a slender-object query selection module is introduced to generate line proposals composed of discrete points. These proposals with spatial cues provide effective initialization for the inputs of the transformer decoder. The CNN-based box head focuses on detecting prevalent small defects and other multi-scale objects. To enhance the detection of complex and large defects, we design a semantic-modulated spatial-coordinated gating module. Through semantic guidance and spatial correlations, this module adaptively gates and weights global context from the line head’s queries to optimize top-level features for the box head. Experiments on an industrial LCD defect detection dataset and two defect datasets of manufacturing domain demonstrate our detector’s superior performance in complex defect detection tasks. Our method outperforms other state-of-the-art methods on the LCD defect dataset, achieving 67.4% mean average precision (mAP) and 88.3% mAP at an intersection over union threshold of 0.5 (mAP 50 ). The code is available at .}, archive = {J_EAAI}, author = {Chaofan Zhou and Meiqin Liu and Senlin Zhang and Shanling Dong and Ronghao Zheng and Shaoyi Du}, doi = {10.1016/j.engappai.2025.112565}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112565}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dual-head detector with point-driven transformer and semantic-spatial gating for liquid crystal display defects}, volume = {162}, year = {2025}, }
• (2025). An integrated machine learning framework for the early diagnosis of hypertension disease. EAAI, 162, 112564. (www)

  @article{ , abstract = {Hypertension is a type of disease that occurs after a serious increase in blood pressure. Due to the rapid increase in the disease, efforts for its early diagnosis are increasing day by day. The use of artificial intelligence (AI) and machine learning (ML) methods in disease detection is of great importance, especially in early diagnosis. In this study, a framework for the diagnosis of hypertension is developed. Behavioral Risk Factor Surveillance System (BRFSS) and Hypertension Risk Prediction (HRP) datasets with clinical and demographic information were used to diagnose hypertension. The problem of class imbalance in the datasets was solved by Random OverSampling method. In particular, the relationship between the attributes and the target variable was analyzed with Consistency-Based Filter, Recursive Feature Elimination with Cross-Validation (RFECV), Least Absolute Shrinkage and Selection Operator (LASSO), Information gain, Relief and Correlation-based feature selection methods and the features selected for the diagnosis of hypertension disease were taken into consideration. In the experiments, Extremely Randomized Trees (Extra Trees), Adaptive Boosting (Adaboost) and Extreme Gradient Boosting (XGBoost) machine learning methods were used in five fold cross-validation. In both datasets, the most successful results were obtained with the XGBoost algorithm. Then, for this algorithm, Random Search, Grid Search, Bayes Search and Genetic Algorithm were used for model hypertuning. For the BRFSS dataset with 87.96 % accuracy and for the HRP dataset with 97.67 % accuracy were obtained. The results are obtained with evaluation metrics that the medical world considers in evaluations, and the proposed framework for hypertension disease detection provides valuable insights for healthcare applications.}, archive = {J_EAAI}, author = {Ayşe Eldem}, doi = {10.1016/j.engappai.2025.112564}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112564}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An integrated machine learning framework for the early diagnosis of hypertension disease}, volume = {162}, year = {2025}, }
• (2025). An explainable deep learning approach for sleep staging in sleep apnea patients across all age subgroups from pulse oximetry signals. EAAI, 162, 112562. (www)

  @article{ , abstract = {Deep-learning (DL) approaches have been developed using pulse rate (PR) and blood oxygen saturation (SpO 2 ) recordings from pulse oximetry to streamline sleep staging, particularly for obstructive sleep apnea (OSA) patients. However, lack of interpretability and validation across patients from a wide range of ages (children, adolescents, adults, and elderly OSA individuals) are two major concerns. In this study, a DL model based on the U-Net framework (POxi-SleepNet) was tailored to accurately perform 4-class sleep staging (wake, light sleep, deep sleep, and rapid-eye movement sleep) in OSA patients across all age subgroups using PR and SpO 2 signals. An explainable artificial intelligence (XAI) methodology based on semantic segmentation via gradient-weighted class activation mapping (Seg-Grad-CAM) was also applied to quantitatively interpret the time and frequency characteristics of pulse oximetry recordings that influence sleep stage classification. Overnight PR and SpO 2 signals from 17303 sleep studies from six datasets encompassing children, adolescents, adults, and elderly OSA individuals were used. POxi-SleepNet showed high performance for sleep staging in the six databases, with accuracies between 81.5 % and 84.5 % and Cohen's kappa values from 0.726 to 0.779. It also demonstrated greater generalizability than previous studies. XAI analysis showed the key contributions of mean and variability in PR and SpO 2 amplitude, as well as changes in their spectral content across specific frequency bands (0.004–0.020 Hz, 0.020–0.100 Hz, and 0.180–0.400 Hz), for sleep stage classification. These findings indicate that POxi-SleepNet could effectively automate sleep staging and assist in diagnosing OSA across all age groups in clinical settings.}, archive = {J_EAAI}, author = {Fernando Vaquerizo-Villar and Gonzalo C. Gutiérrez-Tobal and Daniel Álvarez and Adrián Martín-Montero and David Gozal and Roberto Hornero}, doi = {10.1016/j.engappai.2025.112562}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112562}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An explainable deep learning approach for sleep staging in sleep apnea patients across all age subgroups from pulse oximetry signals}, volume = {162}, year = {2025}, }
• (2025). Transient electrodynamic simulation of laser beam reflection on inclined surfaces using a generative adversarial network-based deep learning model. EAAI, 162, 112561. (www)

  @article{ , abstract = {In laser-material interactions, the temporal dynamics of beam reflection are crucial as they affect both laser absorption and material response. Currently, electrodynamic simulation is the most accurate method for calculating transient laser reflection and absorption patterns, but it is computationally very expensive. In this study, we present the first transient electrodynamic simulation model based on a generative adversarial network, with the time information embedded as color in the input image. The model accurately predicts reflection patterns of a laser beam on inclined surfaces using an input image containing geometric and time information of the domain. The finite-difference time-domain method was used to generate 16 data by changing the surface inclination angle from 0 o to 75 o . The model was based on a generator made by stacking a deep residual network and a discriminator. Although trained on limited data, the model predicted transient beam reflection patterns without overfitting, and the average structural similarity index measure and R-squared accuracy were 96.7 % and 98.0 %, respectively. Another deep learning model was developed to predict the laser beam absorptance from a laser beam reflection image. The ground truth absorptance values were obtained from the Fresnel equations, and the average R-squared accuracy was 99.8 %. Robustness evaluation was additionally performed to examine if the model can be used to predict reflection patterns from angled and curved surfaces. Reasonably accurate results were obtained when the angle difference was not greater than 3° for angled surfaces and the normalized curvature was 0.032 or less for curved surfaces.}, archive = {J_EAAI}, author = {Myeonggyun Son and Hyungson Ki}, doi = {10.1016/j.engappai.2025.112561}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112561}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Transient electrodynamic simulation of laser beam reflection on inclined surfaces using a generative adversarial network-based deep learning model}, volume = {162}, year = {2025}, }
• (2025). Mechanical properties prediction of bi-metal foam sandwiches using machine learning methods and elastic deformation behaviour. EAAI, 162, 112560. (www)

  @article{ , abstract = {Metal foam sandwiches are a kind of ultra-lightweight material made from a porous metal core bonded to two face sheets. Friction stir welding (FSW) is utilised in welding bimetal foam sandwiches. It is worth mentioning that the exact relation between mechanical properties and process parameters is challenging to determine. The innovation lies in the non-destructive estimation of mechanical properties (Young's modulus, ultimate tensile strength and fracture strain) through elastic deformation data and the novel application of artificial intelligence techniques optimised by genetic algorithms, eliminating dependency on input process parameters. After proper network training, three methods are employed to estimate these mechanical properties: a decision tree, a feedforward neural network and long-short term memory. These are chosen to investigate the influence of both machine/deep learning methods in predicting the mechanical properties of the FSW final product. Moreover, a genetic algorithm is employed to find the optimal hyperparameters of the three investigated prediction models to reach the highest accuracy. The results prove the efficiency of the proposed feedforward neural network in the estimation of Young's modulus and ultimate tensile strength for the bi-metal foam sandwiches with lower mean absolute error (MAE) and higher correlation coefficient compared to the decision tree (63.9 % lower MAE and 25.50 % higher correlation coefficient) and long-short term memory (77.50 % lower MAE and 25.05 % higher correlation coefficient). In addition, the proposed decision tree model accurately predicts the fracture strain with R-square and root mean square error as 0.61429 and 1.3862 × 10 −5 , respectively.}, archive = {J_EAAI}, author = {Mohammad Reza Chalak Qazani and Mohsen Dorudgar and Mehdi Moayyedian and Abdel-Hamid I. Mourad and Moosa Sajed and S.M. Hossein Seyedkashi and Siamak Pedrammehr}, doi = {10.1016/j.engappai.2025.112560}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112560}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Mechanical properties prediction of bi-metal foam sandwiches using machine learning methods and elastic deformation behaviour}, volume = {162}, year = {2025}, }
• (2025). Multimodal unified generalization and translation network for intelligent fault diagnosis under dynamic environments. EAAI, 162, 112559. (www)

  @article{ , abstract = {Multimodal data fusion can generate reliable fault representations for intelligent fault diagnosis. However, simple data fusion strategies often introduce fault-irrelevant information, thereby reducing robustness against unknown domain shifts. Moreover, traditional methods generally lack adaptive mechanisms to address missing modalities, leading to considerable performance degradation under sensor failure conditions. To address these problems, this paper proposes a multimodal unified generalization and translation network. To learn invariant unified representations for resisting unknown data distribution shifts, information-enhanced concatenation first generates intra-domain and cross-domain representations. Subsequently, mutual information maximization is applied to remove fault-unrelated information from these representations. Finally, A hybrid ensemble diagnosis strategy fully leverages the interaction of multimodal information across different levels. In addition, semantic supervision investigates the relationships among different modalities and enables intermodal translation in the event of a sensor failure within the monitoring system. Extensive experimental results based on a public bearing dataset and a self-collected motor dataset indicate that the proposed method improves accuracy by 10.53 % and 8.47 % compared to the state-of-the-art methods, respectively. The code and datasets are available at https://github.com/CHAOZHAO-1/MUGTN .}, archive = {J_EAAI}, author = {Chao Zhao and Weiming Shen and Enrico Zio and Hui Ma}, doi = {10.1016/j.engappai.2025.112559}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112559}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multimodal unified generalization and translation network for intelligent fault diagnosis under dynamic environments}, volume = {162}, year = {2025}, }
• (2025). FabricMamba: A fabric surface defect detection system based on large kernel attention and visual state space. EAAI, 162, 112558. (www)

  @article{ , abstract = {In the textile manufacturing industry, fabric defect detection is essential for ensuring product quality. Traditional approaches based on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) often encounter scalability issues, particularly due to the high computational complexity of the self-attention mechanism in ViTs. To address these limitations, this study introduces FabricMamba, a real-time defect detection framework built on the You Only Look Once version 8 (YOLOv8) CNN architecture. The model enhances detection precision and efficiency for complex fabric defects in high-resolution images while minimizing computational cost. YOLOv8 was selected as the base model due to its strong balance between accuracy and inference speed, which is critical in fast-paced textile production settings. FabricMamba extends YOLOv8 with several innovations: the Parallel Large Separable Kernel Attention (P-LSKA) mechanism for multi-scale perception, the Visual State Space Module (MVSS) for long-range dependency modeling, the lightweight DySample module for reduced resource usage, and Programmable Gradient Information (PGI) to optimize training without increasing inference complexity. Extensive evaluations were conducted using a proprietary industrial fabric defect dataset and two public benchmarks, TILDA Textile Texture Database and FDDS Object Detection Dataset. FabricMamba achieved a mean Average Precision (mAP) of 90.0 %, 97.7 %, and 39.1 % on the respective datasets, outperforming the YOLOv8 baseline by 1.8 %, 2.3 %, and 2.0 %. Compared to Mamba-YOLO, FabricMamba reduced model size and computational requirements by 36.7 % and 33.1 %, respectively, with recall improving by 2.9 %, 1.4 %, and 4.0 %. These results confirm the model effectiveness and practical potential for industrial fabric inspection tasks.}, archive = {J_EAAI}, author = {Nengsheng Bao and Jiajun Lin and Yuchen Fan and Runxuan Bao and Alessandro Simeone}, doi = {10.1016/j.engappai.2025.112558}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112558}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {FabricMamba: A fabric surface defect detection system based on large kernel attention and visual state space}, volume = {162}, year = {2025}, }
• (2025). Optimizing performance metrics with new distance and similarity measures using control parameters of linear diophantine fuzzy sets. EAAI, 162, 112557. (www)

  @article{ , abstract = {The control parameters (CPs) of linear Diophantine fuzzy sets (LDFSs) provide an innovative approach for information analysis in the multi-criteria decision making (MCDM), machine learning (ML), and computational intelligence (CI). The CPs give freedom to the decision-makers in evaluating feasible alternatives in measuring performance metrics. Previous fuzzy MCDM methods often meet strict limitations in handling real-world uncertainty and dynamic MCDM. To overcome these limitations, this study extends the Preference Ranking Organization Method for Enrichment of Evaluations (PROMETHEE-II) method to linear Diophantine fuzzy sets (LDFSs) for more flexible and robust MCDM approach. The proposed MCDM approach primarily evaluates performance metrics criterion, utilizing LDFSs which are robust extension of fuzzy sets (FSs), intuitionistic fuzzy sets (IFSs) as well as interval-valued intuitionistic fuzzy sets (IVIFSs). It helps decision makers to address uncertain information with membership degree (MD), non-membership degree (NMD), and CPs. For this objective, new LDFS based distance measures (DMs) and similarity measures (SMs) are developed for the construction of LDFS PROMETHEE-II technique. The proposed MCDM approach provides a structured and comprehensive framework for optimizing investment performance metrics. Its robustness is validated through sensitivity and comparative analyses.}, archive = {J_EAAI}, author = {Masooma Raza Hashmi and Muhammad Riaz and Arshid Mahmood and Muhammad Ajmaeen and Muhammad Aslam}, doi = {10.1016/j.engappai.2025.112557}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112557}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Optimizing performance metrics with new distance and similarity measures using control parameters of linear diophantine fuzzy sets}, volume = {162}, year = {2025}, }
• (2025). A dynamic capsule network-based physical alert monitoring system for elderly people with disabilities. EAAI, 162, 112556. (www)

  @article{ , abstract = {Elderly disabled people have higher health risks due to reduced mobility, delayed emergency response, and poor real-time monitoring. Existing Internet of Things (IoT) based monitoring systems frequently have static thresholds, lack personalization, and struggle to capture complex physical and physiological fluctuations, resulting in false alerts and reduced reliability. We propose a Dynamic Capsule Network-Based Physical Alert Monitoring System (DyCN-PAM) for intelligent, real-time monitoring of elderly disabled people to overcome these constraints. Compiling accelerometer, gyroscope, and heart rate readings, and utilizing capsule routing to preserve spatiotemporal hierarchies, enables the system to detect falls, seizures, and fainting. Context-aware alerts utilize capsule confidence and heart rate baselines to distinguish between typical changes and catastrophic emergencies. The DyCN-PAM system outperforms benchmark models in terms of accuracy and robustness, achieving a 5.13 % increase in F1-score, a 28.6 % increase in precision, a 45.1 % reduction in false alarms, and a 26.3 % improvement in computational efficiency. The DyCN-PAM system enhances accuracy, precision, and efficiency, making it feasible to improve safety, independence, and quality of life for older individuals with disabilities. More real-world experiments are needed to prove its use.}, archive = {J_EAAI}, author = {Shujuan Feng and Hongying Zhu and Yangkai Wu and Ziheng Zeng and Ezzeddine Touti and Jinming Wang and Amar Jain}, doi = {10.1016/j.engappai.2025.112556}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112556}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A dynamic capsule network-based physical alert monitoring system for elderly people with disabilities}, volume = {162}, year = {2025}, }
• (2025). A multiple convolution and bilayer acceleration model for precise and efficient early urban fire detection in complex scenarios. EAAI, 162, 112555. (www)

  @article{ , abstract = {AI advancement enables earlier and more effective urban fire detection, crucial for slowing fire spread. However, hardware limitations make precise and efficient detection under limited resources a major challenge. Moreover, earlier detection of fire requires the identification of smoke, which further exacerbates the difficulty of detecting algorithms since smoke's inherent low-contrast visual properties produce feature blurring from the surrounding background. In this paper, we propose a novel multiple convolutions and bilayer accelerate (MCBA) model for effective early urban fire detection in terms of precision, lightweight and efficiency, which takes advantage of the mainstream You Only Look Once version 8 (YOLOv8) to training and testing the early fire detection model. In our MCBA model, three optimization techniques have been developed to balance lightweight and precision. First, it designs a new multi-convolution (MC) structure to reduce the size of the original backbone network by avoiding complex or skipping connections. Second, the model includes a novel design of a bilayer accelerate mechanism (BAM) at the neck to minimize the interference of redundant background information in multiple scenarios. Third, we provide a precision compensation strategy (PCS) at the neck to enhance the feature extraction and aggregation capabilities, enabling effective detection of small fire areas. The experiments demonstrate that our proposed MCBA model achieves higher performance in terms of precision and efficiency compared with 17 counterpart detection models. It exhibits superior performance with minimal parameter count and the lowest computational complexity among the compared methods. The model shows strong potential for deployment in early urban fire detection across a variety of real-world scenarios.}, archive = {J_EAAI}, author = {Pei Shi and Jun Lu and Yachen Xu and Quan Wang and Yonghong Zhang and Liang Kuang and Deji Chen and Guangyan Huang}, doi = {10.1016/j.engappai.2025.112555}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112555}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multiple convolution and bilayer acceleration model for precise and efficient early urban fire detection in complex scenarios}, volume = {162}, year = {2025}, }
• (2025). Deep learning driven prediction of dynamic stress-strain response in limestone: Insights into transient mechanical behavior under complex loadings for shield tunneling. EAAI, 162, 112554. (www)

  @article{ , abstract = {The complex geological conditions pose a serious challenge to improving the rock-breaking efficiency of shield cutter machines in subway tunnel constructions. This study develops a three-axis thermal-hydraulic-mechanical (THM) coupled dynamic impact system, which is integrated with a scaled shield cutter model to generate a unique dataset of stress wave propagation under multi-field coupling. A bidirectional long short-term memory (LSTM) neural network with an attention mechanism is proposed to establish a nonlinear time mapping relationship between stress waves. The determination coefficient ( R 2 ) exceeds 0.97, the symmetric mean absolute percentage error ( sMAPE ) is less than 10 %, and the average relative uncertainty ( ARU ) is less than 4 %, confirming its high accuracy and reliability. Based on predictions and test results, the new quantitative laws for the transient dynamics of limestone are further revealed. The results reveal that loading conditions do not alter the correlation trend between loading rate and dynamic parameters but significantly influence the degree of the loading rate's effect. Axial pressure dominates energy absorption with an energy contribution rate of 60 %. Confining pressure amplified the sensitivity of loading rate by 170 %, while THM coupling suppressed the dynamic deformation modulus by over 30 %. The accurate prediction of the nonlinear response of stress waves in limestone by the LSTM neural network establishes the connection between transient dynamic observation and rock fragmentation physics mechanism, providing support for quantifying energy absorption and conversion in the rock fragmentation process, and providing key strategies for optimizing shield machine performance in extreme environments.}, archive = {J_EAAI}, author = {Baoping Zou and Kejian Xia and Jingyuan Ma and Xu Long}, doi = {10.1016/j.engappai.2025.112554}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112554}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep learning driven prediction of dynamic stress-strain response in limestone: Insights into transient mechanical behavior under complex loadings for shield tunneling}, volume = {162}, year = {2025}, }
• (2025). A continuous verification mechanism for ensuring client data forgetfulness in federated unlearning. EAAI, 162, 112553. (www)

  @article{ , abstract = {In Federated Learning (FL), it is sometimes necessary to unlearn client data through Federated Unlearning (FU) methods, which help protect user privacy and recover from data poisoning. One critical task is to check the consistency of FU methods and verify if certain clients’ data has been effectively unlearned. However, none of the current FU verification methods can be performed on clients who opt out of the FL process, failing to meet the universal demand in FL, such as the user’s “right to be forgotten” (RTBF). Specifically, after clients leave the FL cooperation, they can no longer verify whether the FL model unlearns their data as the FL continues training for several rounds. To address this, we introduce a continuous verification mechanism for FL clients called Backdoor Attack-based Forgetting Verification (BAFV). The BAFV method embeds a persistent mark for clients who propose to leave, allowing them to verify FU long after leaving the FL cooperation. Extensive experiments across diverse FU environments and datasets demonstrate that our method maintains the model’s accuracy and provides clients with a continuous verification mechanism to ensure their data is unlearned. Our combinatorial marking strategy and gradient-amplified persistence mechanism represent significant advancements beyond existing verification schemes. Our code of BAFV is publicly available at: https://github.com/FuduXing/newBAFV.git .}, archive = {J_EAAI}, author = {Fudu Xing and Jun Liu and Shanshan Chen and Tianlong Yu and Yang Yang}, doi = {10.1016/j.engappai.2025.112553}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112553}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A continuous verification mechanism for ensuring client data forgetfulness in federated unlearning}, volume = {162}, year = {2025}, }
• (2025). Trajectory planning of redundant parallel mechanism considering motion accuracy based on reinforcement learning. EAAI, 162, 112552. (www)

  @article{ , abstract = {The motion accuracy of the trajectory directly affects the reliability of parallel mechanisms in precision tasks such as micro-assembly. Thus, this paper investigates trajectory planning considering motion accuracy, and uses a (6+3)-degrees of freedom (DOF) kinematically redundant parallel mechanism (KRPM) as a case study. First, the kinematics of KRPM are analyzed, and an error model incorporating dimensional errors, driving input errors, and joint clearances is established. Then, based on the error model, the aggregate sensitivity index (ASI) and comprehensive error sensitivity (CES) are introduced to study the error properties of KRPM, along with a universal analysis process. Subsequently, reinforcement learning (RL) utilizing the twin delayed deep deterministic policy gradient (TD3) algorithm is employed to the trajectory planning of KRPM considering motion accuracy. Finally, numerical simulation is carried out based on three cases to verify the effectiveness of the proposed method, and experimental results further demonstrate its practical applicability.}, archive = {J_EAAI}, author = {Chen-dong Zeng and Zhi-cheng Qiu and Fen-hua Zhang and Xian-min Zhang}, doi = {10.1016/j.engappai.2025.112552}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112552}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Trajectory planning of redundant parallel mechanism considering motion accuracy based on reinforcement learning}, volume = {162}, year = {2025}, }
• (2025). Deep reinforcement learning-tuning hierarchical vehicle trajectory tracking framework based on improved kinematic model predictive control. EAAI, 162, 112551. (www)

  @article{ , abstract = {Trajectory tracking is crucial in vehicle control, as it ensures stable driving along a predefined path. This paper proposes a deep reinforcement learning (DRL)-tuning hierarchical trajectory tracking framework, aiming to improve the tracking accuracy of traditional kinematic model predictive control (MPC) methods in uncertain environments. The proposed hierarchical vehicle trajectory tracking framework consists of two layers: the upper layer serves as a compensation layer for the vehicle side-slip angle (VSA), designed using bidirectional long short-term memory (BiLSTM); while the lower layer is the trajectory tracking layer, in which the improved kinematic MPC is enhanced by integrating the twin delayed deep deterministic policy gradient (TD3) algorithm with an external attention (EA) mechanism. The contribution in artificial intelligence is improving the TD3 algorithm with the EA mechanism, enhancing its ability to capture contextual information and improve adaptability. The contribution in engineering applications is implementing the EA-TD3-tuned hierarchical kinematic MPC framework in the field of vehicle trajectory tracking. With 95 % confidence, compared to traditional kinematic MPC controller, the proposed hierarchical vehicle trajectory tracking framework reduces the average lateral error by 33 % (confidence interval, CI: [0.0616, 0.0850]), the average heading angle error by 34 % (CI: [0.01173, 0.0157]), the average yaw rate variation by 31 % (CI: [0.0244, 0.0346]), and the average front wheel steering angle variation by 28 % (CI: [0.0244, 0.0346]).}, archive = {J_EAAI}, author = {Jiankun Peng and Xingyan Liu and Changcheng Wu and Dawei Pi and Jiaxuan Zhou}, doi = {10.1016/j.engappai.2025.112551}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112551}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep reinforcement learning-tuning hierarchical vehicle trajectory tracking framework based on improved kinematic model predictive control}, volume = {162}, year = {2025}, }
• (2025). A novel decomposition-prediction hybrid model improved by dual-channel cross-attention mechanism for short-term wind speed prediction. EAAI, 162, 112550. (www)

  @article{ , abstract = {Accurate prediction of wind energy is crucial for ensuring the safe operation and stability of power systems. To improve the accuracy and robustness of wind speed (WS) forecasting, a novel hybrid method based on mixture of experts (MoE), Transformer, temporal convolution network (TCN), multi-head attention (MA) mechanism, and improved by dual-channel cross-attention mechanism (DCCAM) is proposed. The wind speed data is decomposed by MoE into seasonal and trend components. The trend features are directly captured through a multi-head attention mechanism. An innovative Transformer-TCN framework associated with DCCAM is designed to handle the seasonal component, wherein the Transformer-TCN can make full use of long dependency modeling of Transformer and the local feature extraction of TCN, and DCCAM enables the information interchange and feature fusion, therefore realizing complementary advantages. Based on the proposed model, a series of experiments are conducted on multiple datasets. Ablation experiments confirm the effectiveness of the model and the role of each module in performance improvement. Experiments on seasonal datasets, including spring, summer, autumn, and winter, show that the proposed model can effectively adapt to variations in amplitude and volatility of wind speed sequences under different climatic conditions. Comparative experiments with six advanced hybrid models including decomposition and prediction modules, further demonstrate the superiority and stability of the proposed model.}, archive = {J_EAAI}, author = {Donghan Geng and Haiteng Cui and Leisen Lv and Jiamin Guo}, doi = {10.1016/j.engappai.2025.112550}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112550}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel decomposition-prediction hybrid model improved by dual-channel cross-attention mechanism for short-term wind speed prediction}, volume = {162}, year = {2025}, }
• (2025). Analysing sustainable industrial wastewater treatment technologies using circular fermatean fuzzy multi-attribute group decision making with decision experts’ confidence levels. EAAI, 162, 112549. (www)

  @article{ , abstract = {The evaluation of sustainable industrial wastewater treatment techniques is vital for preserving environmental integrity and protecting public health. Industrial processes often generate wastewater containing toxic compounds like metal contaminants, toxic chemicals, and complex organic compounds, posing serious risks to ecosystems and human well-being. This study proposes a robust multi-attribute group decision-making framework to assess five treatment alternatives across twelve sub-criteria. The evaluation model employs circular Fermatean fuzzy numbers to capture uncertainty and imprecision in expert judgements. To enhance the accuracy of data aggregation, four novel Schweizer–Sklar weighted aggregation operators are introduced, integrating varying confidence levels. Criteria weights are determined through a hybrid approach combining the subjective Ranking Comparison (RANCOM) method and the objective Opinion Weight Criteria Method (OWCM), ensuring balanced prioritization. Alternatives are ranked using the Alternative Ranking Order Method Accounting for Two-Step Normalization (AROMAN), a novel technique for improved discrimination and consistency. Results reveal that the membrane bioreactor as the most sustainable treatment with score 0.821, outperforming activated sludge process, by 25.34%. The lowest-ranked option is chemical coagulation and flocculation, scoring 0.622. Sensitivity analysis, performed by varying three parameters, shows reasonable stability with an average correlation value of 0.71. Comparative analysis shows an average Spearman’s rank correlation of 0.86, confirming reliability. The study recommends prioritizing membrane bioreactor adoption in industrial treatment plants to enhance efficiency and water reuse. By promoting effective treatment solutions, the study contributes to reducing industrial pollution, enhancing water reuse, and advancing environmental sustainability.}, archive = {J_EAAI}, author = {Prayosi Chatterjee and Mijanur Rahaman Seikh}, doi = {10.1016/j.engappai.2025.112549}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112549}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Analysing sustainable industrial wastewater treatment technologies using circular fermatean fuzzy multi-attribute group decision making with decision experts’ confidence levels}, volume = {162}, year = {2025}, }
• (2025). A cost-effective nash-based allocation method for task distribution of multiple robots in distributed robotic networks. EAAI, 162, 112548. (www)

  @article{ , abstract = {Efficiency is paramount in distributed robotic networks (DRNs), where multiple autonomous robots collaborate to perform complex tasks. In this context, the identification of the most efficient path for robots, considering both distance and cost, plays a crucial role in the development of an effective matching algorithm for addressing multirobot task allocation (MRTA) challenges. The study introduces a new cost-efficient Nash-based game framework for task allocation in a distributed robotic network. The proposed model relies on a decentralized decision-making strategy, where each robot selects a single task that optimizes its execution time at a constant speed, thereby maximizing energy harvesting and minimizing energy consumption. In this context, each robot optimizes its choices for individual benefit while also considering the collective welfare, achieving the Nash equilibrium as a nearly optimal allocation strategy in DRNs. The proposed model is tested on various MRTA scenarios involving five robots, seven robots, ten robots, fifteen robots, and twenty robots with the same number of tasks. The proposed Nash-based decentralized model outperforms the Hungarian method by significantly reducing computational costs and complexity to O ( N ) , making it more efficient for large-scale problems.}, archive = {J_EAAI}, author = {Ali Hamidoğlu and Omer Melih Gul and Seifedine Nimer Kadry and Chiranjibe Jana and Ali Elghirani and Gokhan Koray Gultekin}, doi = {10.1016/j.engappai.2025.112548}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112548}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A cost-effective nash-based allocation method for task distribution of multiple robots in distributed robotic networks}, volume = {162}, year = {2025}, }
• (2025). A digital twin-assisted algorithm for diagnosis of permanent magnet synchronous generator interturn short circuit fault and converter open circuit fault in wind power systems using pearson correlation coefficient. EAAI, 162, 112547. (www)

  @article{ , abstract = {Interturn short-circuit faults (ISCFs) in permanent magnet synchronous generators (PMSGs) and open-circuit faults (OCFs) in the machine-side converters represent two critical reliability challenges in wind power systems. Conventional fault diagnosis approaches typically rely on dedicated models for each fault type for each fault type, leading to excessive system complexity and suboptimal computational efficiency. To overcome these limitations, this paper proposes a novel unified digital twin-assisted framework capable of simultaneous diagnosis of both PMSG ISCFs and converter OCFs within a single integrated architecture. The high-fidelity digital twin model based on one-dimensional convolutional neural networks is established to generate real-time reference value of current space vector (SV) for online fault detection, while Pearson correlation coefficient analysis enables accurate differentiation between ISCF and OCF. For ISCFs, the fault severity assessment is performed based on the deviation between reference and measured current SV, with the faulty phase identified using phase current root mean square (RMS) values. In the case of converter OCFs, the proposed method introduces a dual-stage identification process: single and dual insulated-gate bipolar transistor (IGBT) open faults are differentiated through severity estimation analysis, and the faulty IGBT is identified by evaluating the effective current interval ratio (ECIR) and normalized current average (NCA). The experimental results validate the effectiveness of the proposed method, and comparative analysis further demonstrates its superior performance in terms of parameter dependency and diagnostic efficacy.}, archive = {J_EAAI}, author = {Bin Sun and Ying Zhu and Zhinong Wei}, doi = {10.1016/j.engappai.2025.112547}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112547}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A digital twin-assisted algorithm for diagnosis of permanent magnet synchronous generator interturn short circuit fault and converter open circuit fault in wind power systems using pearson correlation coefficient}, volume = {162}, year = {2025}, }
• (2025). A lightweight self-attention metric network for bird species recognition in intelligent bird repellent equipment. EAAI, 162, 112546. (www)

  @article{ , abstract = {Bird damages to power transmission lines pose significant operational risks, and intelligent bird repellent equipment (IBRE) requires accurate species recognition for effective long-term repellent. We propose a novel lightweight self-attention metric network (LSAM-Net) for few-shot bird species recognition in the vicinity of power transmission lines, aiming to enhance the performance of IBRE. LSAM-Net integrates a simple attention mechanism (SimAM) to emphasize critical spatial and channel features, thereby enhancing the extraction of key semantic information from bird images. Additionally, a self-correlation representation (SCR) module is employed to capture local structural patterns, effectively mitigating the impact of pseudo-features and improving the network’s capacity to learn discriminative representations. To promote the utilization of local discriminative information in few-shot classification, LSAM-Net leverages earth mover’s distance (EMD) to compute structural similarity between images. For efficient deployment, we apply knowledge distillation to further reduce model complexity. Extensive experiments conducted on Bird-65, CUB200, 2011, miniImageNet, and Fewshot-CIFAR100 demonstrate that LSAM-Net achieves superior performance compared to state-of-the-art methods, while maintaining a compact architecture. On the Bird-65 and CUB200-2011 datasets, LSAM-Net requires only 4.75 and 1.18 giga floating-point operations (GFLOPs), and achieves inference speed improvements of 52.9 % and 48.9 %, respectively, over the self-attention metric network (SAM-Net). Further optimization with TensorRT yields additional reductions in inference time by 43.6 ms and 53.7 ms, respectively. These improvements significantly support species-specific repellent strategies, thereby enhancing the long-term effectiveness of IBRE systems.}, archive = {J_EAAI}, author = {Jiangjian Xie and Shanshan Xie and Baican Li and Yujie Zhong and Chunhe Hu and Junguo Zhang and Björn W. Schuller}, doi = {10.1016/j.engappai.2025.112546}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112546}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight self-attention metric network for bird species recognition in intelligent bird repellent equipment}, volume = {162}, year = {2025}, }
• (2025). A multimodal multi-scale fusion network for leak detection in marine piping systems. EAAI, 162, 112545. (www)

  @article{ , abstract = {Marine system monitoring data inherently exhibit multimodal characteristics, making artificial intelligence-driven correlation and fusion essential for improving fault feature recognition. However, existing intelligent diagnosis methods mostly focus on feature fusion within homogeneous data types, such as fusing multiple time-series signals or multiple image sets, while systematic exploration of joint representation learning across heterogeneous dimensions remains under-explored. This limitation constrains the recognition capability for complex failure modes. Meanwhile, the inherent differences in physical meanings and representations of multimodal data pose significant challenges in constructing effective correlations, often limiting the performance of mainstream machine learning based fault diagnosis approaches. The proposed method enhances the fault diagnosis capability of mainstream approaches through the fusion of multi-sensor data and visual data, with its core innovation residing in a multimodal fusion framework leveraging attention mechanisms to effectively integrate cross-dimensional representations of multivariate time-series data and imaging data. Compared to existing multimodal transformer techniques, this dual-strategy architecture enables the model to simultaneously capture shared systemic behaviors and modality-unique signatures, substantially elevating diagnosis precision. Experimental validation on real-world leak detection datasets demonstrates that the proposed model achieves F1-scores consistently surpassing 90 % across diverse marine monitoring scenarios, with quantitative evaluations further confirming its superior performance over conventional multivariate time-series diagnosis methods in establishing multimodal correlations, conclusively validating both technical excellence and engineering practicability.}, archive = {J_EAAI}, author = {Peng Zhang and Chaozhe Li and Shitao Peng and Bomu Tian and Si Luo and Yuewen Zhang and Taili Du}, doi = {10.1016/j.engappai.2025.112545}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112545}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multimodal multi-scale fusion network for leak detection in marine piping systems}, volume = {162}, year = {2025}, }
• (2025). A unified rotating machinery health management framework leveraging large language models for diverse components, conditions, and tasks. EAAI, 162, 112544. (www)

  @article{ , abstract = {This study introduces the Rotating Machinery Large Language Model (RotLLM), a unified framework for rotating machinery health management that integrates deep learning with large language models (LLMs) to address diverse operational conditions, components, and health management tasks. RotLLM employs a novel Spectral Folding Network (SFN) to transform vibration spectrum into a unified feature space that preserves essential health state information. A dedicated projection layer then maps these features into the semantic domain of an LLM. The framework is trained using a three-stage strategy: first, pre-training the encoder on the Large-scale Multimodal Rotating Machinery (LMR) dataset, which comprises 237,298 vibration samples collected under hundreds of operating conditions; second, initializing the projection layer with textual health state labels; and finally, fine-tuning using parameter-efficient Low-Rank Adaptation (LoRA) with high-quality corpus for various health management tasks. Experimental evaluations demonstrate that RotLLM achieves state-of-the-art performance in fault classification, maintains strong robustness under noisy conditions, and delivers rapid multi-task inference with minimal computational overhead. The framework consistently outperforms conventional methods, enabling efficient, accurate, and context-aware health management for rotating machinery across diverse conditions and tasks. The dataset and source code are open-sourced ( https://github.com/SIA-IDE/RotLLM ), fostering collaboration, reproducibility, and broader adoption in industrial prognostics research.}, archive = {J_EAAI}, author = {Haotian Peng and Jie Gao and Jiawei Liu and Jinsong Du and Wei Wang}, doi = {10.1016/j.engappai.2025.112544}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112544}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A unified rotating machinery health management framework leveraging large language models for diverse components, conditions, and tasks}, volume = {162}, year = {2025}, }
• (2025). A hybrid neural network model for shale gas production prediction considering production plans and produced water. EAAI, 162, 112543. (www)

  @article{ , abstract = {Recurrent neural networks have been widely applied to shale gas production prediction. However, current literature has two major flaws. Firstly, the delay issue can be observed in single-step prediction. Secondly, models that can be directly applied to newly developed wells are underresearched. This study presents a hybrid neural network combining fully connected (Dense) layers and long short-term memory (LSTM) layers to solve the two problems. In the proposed Dense-LSTM model, parameters regarding produced water and production plans are fed into the Dense layers while historical data is fed into the LSTM layers. The bottom-hole total flow rate converted from ground gas and water rates using volume factors is also included as an auxiliary parameter. The production plans made from the feature engineering can preserve the key information from shut-in periods while keeping a continuous prediction. The performance of the proposed model is demonstrated through ablation experiments and the comparison with popular decline curve models. For long-term prediction with window size w = 14 days, the proposed model’s mean cumulative error is 40.26% of that of vanilla LSTM at the end of the 3-year-long prediction interval. For window size w ′ = 100 days, the proposed model’s mean cumulative error is 77.93% of that of vanilla LSTM and 29.71% of that of the classic Arps model. The workflow of this study also sheds light on the construction of a universal prediction model for shale gas wells from different gas fields.}, archive = {J_EAAI}, author = {Yilun Dong and Youzhi Hao and Detang Lu}, doi = {10.1016/j.engappai.2025.112543}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112543}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A hybrid neural network model for shale gas production prediction considering production plans and produced water}, volume = {162}, year = {2025}, }
• (2025). A novel multi-source heterogeneous data fusion based fault diagnosis framework for manufacturing processes. EAAI, 162, 112542. (www)

  @article{ , abstract = {Fault diagnosis technologies are important means to ensure the production operation safety and product quality stability for manufacturing processes. After a fault occurs in the manufacturing processes, it may be characterized by high-frequency and high-dimensional structured time series data anomalies such as sensor data, or by unstructured data anomalies such as images. Traditionally, single structured sensor data is often used for constructing diagnosis models, the fault characteristics may not be adequately characterized, thus affecting the diagnosis performance. Therefore, in this paper, in order to make full use of multi-source data and obtain more comprehensive and accurate diagnosis results, a new multi-source heterogeneous data fusion based fault diagnosis framework is designed for manufacturing processes. Specifically, to solve the problem that the important information is ignored during data level fusion of multi-source data, an adaptive weight multi-source data fusion method is proposed. Furthermore, in response to the problem of feature redundancy in feature level fusion of heterogeneous data, a feature differentiation extraction and heterogeneous feature fusion method is proposed, of which a feature source discriminator is constructed for enhancing the complementarity of the extracted heterogeneous features, and feature concatenation is performed to improve the feature expression ability. Finally, the effectiveness and feasibility of the proposed framework is verified on actual datasets from the hot rolling process and the Tennessee Eastman process. Experimental results show that the proposed framework is both effective and feasible in fault diagnosis with multi-source heterogeneous data.}, archive = {J_EAAI}, author = {Liang Ma and Qikai Yang and Orestes Llanes-Santiago and Kaixiang Peng}, doi = {10.1016/j.engappai.2025.112542}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112542}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel multi-source heterogeneous data fusion based fault diagnosis framework for manufacturing processes}, volume = {162}, year = {2025}, }
• (2025). Reducing modal differences in zero-shot anomaly detection based on vision-language generation model. EAAI, 162, 112541. (www)

  @article{ , abstract = {Zero-shot anomaly detection methods based on vision-language model rely on alignment between image and text. These methods ignore the inherent differences between different modalities, which is unfavorable for improving the alignment between modalities. This paper reduces modal differences between image and text by using guiding vision feature and text feature from the pre-trained vision-language generation model. The vision perception text embedding is constructed by adding guiding vision feature to the weight shared text prompt. The text perception vision embedding is extracted by a vision text fusion module. The fusion module is designed to promote the visual modality to perceive the textual information locally. Anomaly regions are detected by cosine similarity between cross-modal perception embeddings. Zero-shot anomaly detection performance is evaluated on five publicly available industrial anomaly detection datasets, and a real-world dataset about automotive plastic parts. Experimental results show that the proposed method achieves highly competitive anomaly detection performance on multiple evaluation metrics.}, archive = {J_EAAI}, author = {Yanan Song and Weiming Shen and Baisong Pan and Quanhui Wu and Dawei Gu}, doi = {10.1016/j.engappai.2025.112541}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112541}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Reducing modal differences in zero-shot anomaly detection based on vision-language generation model}, volume = {162}, year = {2025}, }
• (2025). A novel short-term prediction method for distributed photovoltaic power generation considering extreme weather. EAAI, 162, 112540. (www)

  @article{ , abstract = {Distributed photovoltaic power plants are often impacted by various factors such as weather conditions and geographical locations, making it challenging to fully capture the spatial correlation characteristics among multiple photovoltaic plants. Furthermore, the failure to consider meteorological factors that influence photovoltaic output results in larger prediction errors during extreme weather events. To reduce prediction errors, this paper proposes a short-term photovoltaic forecasting method that considers meteorological factors, explores spatial correlations among photovoltaic plants, and captures temporal characteristics. Firstly, a Graph Attention Network is established to obtain spatial correlations between different plants while a Convolutional Neural Network is employed to extract feature information of meteorological factors. Then, the feature information from these two sources is integrated and input into a Long Short-Term Memory network, which is enhanced based on Spiking Neural P Systems to extract temporal characteristics of photovoltaic output and complete the prediction task. Finally, real-world power station datasets are utilized for validation and comparison with several typical photovoltaic prediction models. The results clearly show that the application of artificial intelligence in this proposed method can effectively improve the accuracy of distributed photovoltaic power forecasting, demonstrating the great potential of AI in the field of photovoltaic power prediction.}, archive = {J_EAAI}, author = {Xin Guan and Xiao Han and Jun Wang and Tao Wang}, doi = {10.1016/j.engappai.2025.112540}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112540}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel short-term prediction method for distributed photovoltaic power generation considering extreme weather}, volume = {162}, year = {2025}, }
• (2025). An innovative method of multi-view face frontalization based on receptive field-enhanced conditional generative adversarial network. EAAI, 162, 112539. (www)

  @article{ , abstract = {A series of theoretical and experimental investigations based on deep learning were conducted to enhance the accuracy and generalizability of face frontalization models. A series of theoretical investigations and experimental verifications based on deep learning are conducted to further enhance the face frontalization model's accuracy and generalizability. To address inadequate local feature extraction and the low realism in synthesized images, a receptive field-enhanced conditional generative adversarial network (RFC-GAN) is proposed to achieve multi-view face frontalization. RFC-GAN model integrates a novel configuration of multi-scale dilated convolutions in a multi-branch generator architecture to significantly expand the receptive field and improve feature extraction. The unique integration enhances the realism and detail of the generated images. Unlike conventional approaches that focus primarily on pixel-level accuracy, RFC-GAN introduces a perceptual loss component to enhance semantic content and structural integrity at the feature level. RFC-GAN has been experimentally validated on the Karolinska Directed Emotional Faces (KDEF) and Carnegie Mellon University Multiple Pose, Illumination, and Expression Face Database (CMU Multi-PIE). The generated facial expression images from RFC-GAN exhibit a higher degree of detailed texture reproduction in critical facial features such as the eyes, nose, and mouth. On the two datasets, the Peak Signal-to-Noise Ratio (PSNR) reaches 31.5185 for KDEF and 26.1851 for Multi-PIE, the Structural Similarity Index (SSIM) reaches 0.3604 for KDEF and 0.3965 for Multi-PIE, and the Learned Perceptual Image Patch Similarity (LPIPS) reaches 0.117 for KDEF and 0.408 for Multi-PIE, respectively. Compared to existing state-of-the-art methods, RFC-GAN exhibits marked improvements in these metrics, especially in detailed texture reproduction of critical facial features such as the eyes, nose, and mouth, establishing new benchmarks in face frontalization.}, archive = {J_EAAI}, author = {Yancong Zhou and Dongdong Wang}, doi = {10.1016/j.engappai.2025.112539}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112539}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An innovative method of multi-view face frontalization based on receptive field-enhanced conditional generative adversarial network}, volume = {162}, year = {2025}, }
• (2025). Enhanced air pollution spatiotemporal forecast model using frequency domain convolution and attention mechanism. EAAI, 162, 112538. (www)

  @article{ , abstract = {Accurate forecasting of air pollution, which is crucial for public health and environmental management, often faces challenges in effectively capturing the complex and intertwined spatiotemporal dynamics of pollutants. Existing models frequently struggle to simultaneously account for broad periodic spatiotemporal dependencies as well as fine-grained local temporal patterns. This paper presents a novel deep learning architecture, the Fourier Convolutional Graph Transformer (FCGformer), specifically designed to overcome these limitations. FCGformer distinctively features a dual-module approach: a Global Module that constructs an integrated spatiotemporal graph and leverages Fourier transforms with frequency domain convolution to extract long-range dependencies and crucial periodicities; and a Local Module that employs inverse temporal embedding and self-attention to meticulously capture nuanced, short-term temporal variations. The key contribution of this work lies in the synergistic integration that enables FCGformer to effectively model complex pollutant behaviors, providing a more comprehensive understanding of both global contexts and local details. Extensive experiments demonstrate that FCGformer significantly outperforms state-of-the-art benchmark models in prediction accuracy, offering a promising advancement for improved air quality management.}, archive = {J_EAAI}, author = {Haiwei Yang and Ru Yang and Ling Ding and Shiqiang Du and Maozhen Li and Bo Zhang}, doi = {10.1016/j.engappai.2025.112538}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112538}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhanced air pollution spatiotemporal forecast model using frequency domain convolution and attention mechanism}, volume = {162}, year = {2025}, }
• (2025). Gate-guided spatial-channel reconstruction network: An efficient lightweight framework for steel surface defect detection. EAAI, 162, 112537. (www)

  @article{ , abstract = {To address the limitations in You Only Look Once version 8 (YOLOv8) for steel surface defect detection, including insufficient generalization of image enhancement, constrained feature representation capability in core modules, and poor adaptability of the loss function to scale variations and sample imbalance, this paper proposes the Gate-guided Spatial-channel Reconstruction Network, an efficient and lightweight improved network. Contrast-Limited Adaptive Histogram Equalization (CLAHE) is introduced to enhance local image details and contrast while reducing noise impact. The Gating Block and the Spatial-channel Reconstruction Block are designed to replace the original C2f (cross-stage partial bottleneck with two convolutions) module in YOLOv8, thereby enhancing feature representation capability and efficiency. The loss function is optimized using Wise-IoU (WIoU) and Slide Loss (SlideLoss) to improve convergence and robustness. The proposed network was evaluated on the Northeastern University Surface Defect Detection (NEU-DET) dataset (200 × 200 pixels) and the Chinese Academy of Sciences Defect Detection (GC10-DET) dataset (2048 × 1000 pixels). It demonstrated high detection accuracy, achieving the mean Average Precision at 50 % (mAP50) of 84.7 % and 79.4 %, respectively. Furthermore, the network maintains low complexity with only 3.6 million parameters and achieves a high detection speed of up to 154 frames per second (FPS). The Gate-guided Spatial-channel Reconstruction Network effectively detects surface defects on hot-rolled steel, achieving state-of-the-art detection accuracy. It successfully meets the requirements for precise and real-time steel surface defect detection under resource-constrained industrial conditions.}, archive = {J_EAAI}, author = {Wei Zhang}, doi = {10.1016/j.engappai.2025.112537}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112537}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Gate-guided spatial-channel reconstruction network: An efficient lightweight framework for steel surface defect detection}, volume = {162}, year = {2025}, }
• (2025). Tensor product-fault diagnosis-transformer based wind turbine blade fault prediction method. EAAI, 162, 112535. (www)

  @article{ , abstract = {With the continuous iterative updating of wind turbine (WT) blade fault diagnosis (FD) technology, intelligent prediction methods based on supervisory control and data acquisition (SCADA) systems have gradually become advanced mainstream technology in the industry. However, despite the many advantages of SCADA data in fault prediction, its high-dimensional characteristics and highly unstable nature still pose significant challenges for practical applications. Accordingly, this study proposes an innovative WT blade fault prediction method based on tensor product dimensionality reduction and FD-Transformer (TP-FD-Transformer) methodology, which aims to effectively solve the problem of timely and accurate prediction of WT blade faults. The TP-FD-Transformer method combines the quantum dimensionality reduction technique with the FD-Transformer model to form a new framework for data processing and analysis. The TP-FD-Transformer method adopts the tensor product-relative position matrix composite dimensionality reduction technique, which effectively reduces the dimensionality and complexity of SCADA data while preserving its features. After data processing is completed, the TP-FD-Transformer method utilizes the FD-Transformer model for deep learning training. The FD-Transformer model has been improved for complex time series data and can effectively capture potential features in the data. The experiments under the open dataset show that the TP-FD-Transformer method demonstrates excellent prediction ability in the field of WT blade FD, with an accuracy rate of 93.65 %. The research findings verify that TP-FD-Transformer method provides a feasible solution for the intelligent diagnosis of WT blade faults, with broad application prospects and significant practical significance.}, archive = {J_EAAI}, author = {Linfei Yin and Yuhan Liu and Nannan Wang}, doi = {10.1016/j.engappai.2025.112535}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112535}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Tensor product-fault diagnosis-transformer based wind turbine blade fault prediction method}, volume = {162}, year = {2025}, }
• (2025). Hybrid physics-based and data-driven method for the rotor angle prediction. EAAI, 162, 112533. (www)

  @article{ , abstract = {Deep learning methods are widely employed for transient stability analysis in power systems. However, the prediction performance can be adversely affected by undesirable factors in the measured data. To alleviate the effect of the undesirable factors, a hybrid physics-based and data-driven prediction network of the rotor angle trajectory is proposed in this paper. The prediction model embeds the rotor equation to form a dynamic learning model that conforms to the actual physical law. The physical consistency of the prediction results is guaranteed. Meanwhile, a dynamic error derivative integral network incorporating the Runge–Kutta method is proposed to correct the final results. The accuracy of the prediction can be improved. Finally, it is tested in the IEEE 39-bus system and the East China Power Grid system. The test results show that the model significantly outperforms other comparative models. And the dependence on the quality of measured data can be alleviated effectively.}, archive = {J_EAAI}, author = {Lingzhe Zhang and Dong Huang and Huaiyuan Wang}, doi = {10.1016/j.engappai.2025.112533}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112533}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hybrid physics-based and data-driven method for the rotor angle prediction}, volume = {162}, year = {2025}, }
• (2025). Physics-informed machine learning for near real-time stress prediction on a structural component: Application for landing gears. EAAI, 162, 112532. (www)

  @article{ , abstract = {Lightweight design constitutes a pivotal research and development objective for next-generation landing gear systems. Nevertheless, achieving reduced weight while maintaining structural safety and reliability presents considerable challenges. The establishment of a digital twin (DT) for structural health monitoring (SHM) offers a promising approach to address these concerns across the design, testing, and operational lifecycle of landing gears. In this study, we develop a physics-informed neural network (PINN) model for near real-time stress prediction on the drag strut of a nose landing gear (NLG), specifically for an A320-type aircraft, serving as a foundational component of a DT system. The proposed PINN framework directly outputs displacement fields while deriving stresses as secondary quantities, effectively incorporating the fundamental equations of linear elasticity into the loss function. Displacement boundary conditions, informed by finite element method (FEM) simulations, are integrated as penalty terms to enhance trainability and physical consistency. The training dataset is constructed using load cases statistically representative of actual landing gear operations, with high-fidelity FEM providing corresponding displacement and stress references. The model demonstrates strong predictive accuracy, with relative errors between 5% and 7% compared to FEM results, and significantly outperforms both pure stress-output PINNs and conventional deep neural networks (DNNs). Moreover, the trained PINN achieves inference times within seconds under time-varying loads, highlighting its capability for near real-time stress monitoring. This work underscores the potential of physics-informed machine learning for enhancing DT-enabled SHM systems in safety-critical aerospace structures.}, archive = {J_EAAI}, author = {Zixuan Zhu and Yifan Zhao and Agusmian Partogi Ompusunggu}, doi = {10.1016/j.engappai.2025.112532}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112532}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Physics-informed machine learning for near real-time stress prediction on a structural component: Application for landing gears}, volume = {162}, year = {2025}, }
• (2025). Enhanced forest fire detection via dynamic multiscale fusion and contextual partial cross features. EAAI, 162, 112531. (www)

  @article{ , abstract = {Timely and accurate detection of forest fires, particularly in the early stages when smoke and small flames are present, is crucial for minimizing ecological damage and improving the effectiveness of emergency response. However, existing methods face challenges such as missing edge information, ineffective multi-scale feature fusion, and low accuracy when identifying small or distant targets in complex forest conditions. To address these issues, a novel detection framework is proposed, called the Dynamic Contextual Shallow Network (DCSNet). This framework enhances detection performance and real-time efficiency. The proposed method incorporates three key components: (1) the Contextual Partial Cross Feature Network (CPCFNet), which employs a reparameterized non-local attention mechanism and partial channel separation to strengthen contextual representation; (2) the Dynamic Multiscale Fusion Pyramid Network (DMFPN), which uses dynamic sampling and deformable convolution to fuse multi-scale features adaptively; and (3) the Shallow Feature Detection Layer (SFDL), which refines shallow features to improve the detection of small smoke and flame targets. Experimental evaluations on visible-light and infrared remote sensing datasets collected by unmanned aerial vehicles (UAVs) demonstrate the effectiveness of DCSNet. Specifically, DCSNet achieves an average mean average precision (mAP) at 50 % intersection over union (IoU) of 76.0 %, a frame rate of 283 frames per second (FPS), and an F1-score of 71.9 % on the visible-light dataset. On the infrared dataset, the framework achieves an mAP@50 of 77.5 %, an mAP@50:95 of 45.5 %, and an F1 score of 72.9 %. These results suggest that DCSNet provides high accuracy and robustness in the real-time detection of forest fires under diverse environmental conditions. The code and datasets are available at https://github.com/Lili-wang-del/DCSNet .}, archive = {J_EAAI}, author = {Lili Wang and Lei Guo and Haiyan Li and Bingbing He and Jundong Yang and Yaqun Huang}, doi = {10.1016/j.engappai.2025.112531}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112531}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhanced forest fire detection via dynamic multiscale fusion and contextual partial cross features}, volume = {162}, year = {2025}, }
• (2025). Enhanced targeted attacks on graph neural networks via average gradient and perturbation optimization. EAAI, 162, 112530. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) are vulnerable to adversarial attacks that cause performance degradation by adding small perturbations to the graph. Gradient-based attacks are among the most widely used methods and have demonstrated strong performance across various attack scenarios. However, most gradient attacks use greedy strategies to generate perturbations, which tend to fall into local optima, leading to underperformance of the attack. To address the above problem, we propose an attack (Average Gradient and Perturbation Optimization Attack, AGPOA) on GNNs, which consists of an average gradient calculation and a perturbation optimization module. In the average gradient calculation module, we compute the average of the gradient information over all moments to guide the attack to generate perturbed edges, which stabilizes the direction of the attack update and gets rid of undesirable local maxima. We use a perturbation optimization module to limit the attack budget and further improve performance. Furthermore, we demonstrate the theoretical superiority of AGPOA over traditional gradient-based attack methods through attack loss variance. The experimental results show that AGPOA improves the misclassification rate by 2%–8% compared to other state-of-the-art models in the node classification task.}, archive = {J_EAAI}, author = {Yang Chen and Bin Zhou and Haixing Zhao and Padarti Vijaya Kumar}, doi = {10.1016/j.engappai.2025.112530}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112530}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhanced targeted attacks on graph neural networks via average gradient and perturbation optimization}, volume = {162}, year = {2025}, }
• (2025). Model-free safe deep reinforcement learning for grid-to-vehicle management considering grid constraints and transformer thermal stress. EAAI, 162, 112529. (www)

  @article{ , abstract = {The increasing penetration of Electric Vehicles (EVs) presents challenges to the distribution grid, due to more volatile power profiles and higher peak demand. One key research question is how to accommodate EVs with limited-capacity grid equipment, such as transformers and lines. However, uncertainties from the EV side and the complexity of grid equipment models challenge the performance of the control strategies implemented. Moreover, the thermal loading of the transformer is often neglected. In this work, we propose a fully model-free, safe Deep Reinforcement Learning (DRL)- based grid-to-vehicle management strategy to avoid electric and thermal overloading of the transformer and power grid constraint violation. The management strategy is based on Projection-based Constraint Policy Optimization (PCPO) and takes only the observable information from the grid and vehicles. The target is to maximize energy delivery to the EV fleet while considering safe constraints, such as transformer thermal loading, voltage magnitude limits, and line loading limits. We compared the proposed strategy with conventional DRL and other safe DRL methods and investigated its robustness against higher ambient temperatures. The results show that the proposed strategy can deliver 92 % energy and reduce violations of the grid and transformers, while the other benchmarks deliver less than 80 %. The robustness test demonstrates that the proposed strategy is effective in various temperature. Moreover, the proposed strategy can effectively reduce at most 90 % of the transformer aging incurred by the thermal stress, compared with the uncontrolled charging.}, archive = {J_EAAI}, author = {Zhewei Zhang and Rémy Rigo-Mariani and Nouredine Hadjsaid and Yan Xu}, doi = {10.1016/j.engappai.2025.112529}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112529}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Model-free safe deep reinforcement learning for grid-to-vehicle management considering grid constraints and transformer thermal stress}, volume = {162}, year = {2025}, }
• (2025). Probability forecasting for multivariate urban water demand using temporal convolutional network based on quantile regression and parzen window. EAAI, 162, 112528. (www)

  @article{ , abstract = {Probability prediction can provide more abundant information about uncertainties in future water demand, which is gaining increasing attention in building economical and reliable water resource management plans. However, most existing literature on water demand prediction focus on provide deterministic point prediction results. To overcome this problem, a novel hybrid probability forecasting model based on quantile regression temporal convolutional network and Parzen window is proposed for the probability density forecast of multivariate urban water demand. Firstly, to address the complex coupling relationship between water demand and multiple influencing factors, a random forest-based feature selection method is employed to eliminate the redundant variables. Then, a discrete wavelet transform is deployed to decompose the original series into a variety of characteristic subseries to reduce fluctuations of the original water demand series. Secondly, a quantile regression-based temporal convolutional neural network is employed to obtain the conditional quantiles of future water demand. Moreover, a probability density prediction method based on Parzen window estimation is developed to further obtain the distribution information of prediction uncertainty. Finally, a real-world multivariate dataset from a water plant in Suzhou, China, is used for comparison experiments with state-of-the-art models. The comparison results show that the proposed model has achieved an average improvement of 15.4 % and 53.3 % in interval prediction and probability density prediction, respectively. It shows that the proposed model is a reliable prediction model that can assist policymakers to optimize the management of urban water demand.}, archive = {J_EAAI}, author = {Jun Guo and Qingya Meng and Baigang Du and Hui Sun}, doi = {10.1016/j.engappai.2025.112528}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112528}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Probability forecasting for multivariate urban water demand using temporal convolutional network based on quantile regression and parzen window}, volume = {162}, year = {2025}, }
• (2025). Complex T-spherical fuzzy yager prioritized weighted aggregation for prioritizing waste management strategies: A case study from delhi. EAAI, 162, 112527. (www)

  @article{ , abstract = {Efficient waste management is crucial for addressing urban challenges in densely populated areas like Delhi. This study is motivated by a key research gap: the need for an advanced Multi-Criteria Decision-Making (MCDM) framework capable of handling both two-dimensional uncertainty and the prioritized nature of criteria in such complex problems. Our primary contribution is a novel framework integrating the flexibility of Complex T-Spherical Fuzzy Sets (CT-SFS) with Yager’s prioritized aggregation. Unlike traditional T-SFS, which are limited to a one-dimensional representation, CT-SFS incorporates both amplitude and phase information for Membership Degree (MD), Neutral Degree (ND), and Non-membership Degree (NMD), offering a nuanced model for expert evaluations. The framework introduces two innovative operators: the Complex T-Spherical Fuzzy Yager Prioritized Weighted Average (CT-SFYPWA) and the Complex T-Spherical Fuzzy Yager Prioritized Weighted Geometric (CT-SFYPWG). These operators uniquely incorporate Yager’s prioritized aggregation, allowing decision-makers to assign hierarchical importance to critical criteria, which is vital for effective resource allocation in urban waste management. The framework’s practical value is demonstrated through a detailed case study prioritizing waste management strategies in Delhi. Rigorous numerical analysis, sensitivity tests, and comparative evaluations confirm the framework’s robustness, adaptability, and distinct advantages over existing methods. Applying this framework to Delhi, we found that a holistic, integrated waste management system is the most promising strategy, outperforming capital-intensive projects like Waste-to-Energy. Our work offers a powerful, structured methodology for cities to make more nuanced, forward-looking decisions and advances MCDM theory by integrating prioritized aggregation within the CT-SFS environment.}, archive = {J_EAAI}, author = {Fathima Banu M. and S. Solaiappan and Subramanian Petchimuthu and Abrar Hussain}, doi = {10.1016/j.engappai.2025.112527}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112527}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Complex T-spherical fuzzy yager prioritized weighted aggregation for prioritizing waste management strategies: A case study from delhi}, volume = {162}, year = {2025}, }
• (2025). Critical nodes detection for complex networks via knowledge-guided evolutionary framework. EAAI, 162, 112526. (www)

  @article{ , abstract = {The Critical Node Problem (CNP) focuses on identifying critical nodes within complex networks. These nodes play a crucial role in maintaining connectivity, and their removal impacts network performance. Among CNP variants, CNP-1a — which minimizes pairwise connectivity after removing a limited number of nodes — has attracted significant research attention due to its NP-hard nature and applications in diverse fields like epidemic control and infrastructure resilience. While state-of-the-art methods leverage memetic algorithms and variable populations, they fundamentally rely on random initialization that often converges to local optima. This limitation arises because traditional methods fail to capture higher-order topological dependencies. To address this gap, we propose K2GA, a knowledge-guided genetic algorithm initialized by a graph attention network (GAT). The GAT embeds networks into low-dimensional spaces, assigning topology-aware attention weights to nodes that guide population initialization. K2GA then employs a hybrid genetic algorithm with a local search process to identify an optimal set of critical nodes. The local search process utilizes a cut node-based greedy strategy. Experiments on 26 real-world networks demonstrate that K2GA outperforms state-of-the-art methods in terms of the best, median, and average objective values, establishing new upper bounds for minimization in eight cases. This work pioneers a GAT-guided evolutionary search framework, offering a novel paradigm for solving CNP.}, archive = {J_EAAI}, author = {Chanjuan Liu and Shike Ge and Zhihan Chen and Wenbin Pei and Enqiang Zhu and Hisao Ishibuchi}, doi = {10.1016/j.engappai.2025.112526}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112526}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Critical nodes detection for complex networks via knowledge-guided evolutionary framework}, volume = {162}, year = {2025}, }
• (2025). An efficient framework for general long-horizon time series forecasting with mamba and diffusion probabilistic models. EAAI, 162, 112525. (www)

  @article{ , abstract = {Time series forecasting plays an essential role in supporting critical decision-making processes in risk management and resource allocation in various fields, including finance, transportation, industrial systems, etc. Conventional models can effectively capture volatility and, are proficient in handling specific patterns, such as the AutoRegressive Integrated Moving Average model (ARIMA) and the Generalized AutoRegressive Conditional Heteroskedasticity model (GARCH). Nonetheless, these models meet many challenges, such as high dimensionality, non-stationarity, and nonlinearity inherent in real-world data. Although deep learning methodologies can provide better performance, they may still suffer from long-term errors and heightened computational expenses. A novel framework named Mamba Diffusion Probabilistic Models (MambaDiffTS) is proposed, which integrates Mamba’s state space model with a frequency-aware diffusion process grounded in Denoising Diffusion Probabilistic Models (DDPM). Mamba’s selective state transitions enable linear-time modeling of long-range dependencies; at the same time, frequency-aware spectral decomposition isolates trends and seasonality through Fourier regularization. Furthermore, the implementation of spectral energy-guided noise scheduling preserves temporal fidelity. Extensive experiments on diverse benchmarks-financial volatility, industrial IoT sensor data, and climate modeling-demonstrate MambaDiffTS’s superiority. Notably, on stock forecasting tasks, MambaDiffTS reduces Mean Squared Error (MSE) by approximately 18.6% compared to the best-performing baseline, and substantially outperforms diffusion models, all while maintaining linear computational complexity. The proposed MambaDiffTS facilitates scalable forecasting over extended horizons.}, archive = {J_EAAI}, author = {Wenjing Wang and Qilei Li and Ziwu Jiang and Deqian Fu and David Camacho}, doi = {10.1016/j.engappai.2025.112525}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112525}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An efficient framework for general long-horizon time series forecasting with mamba and diffusion probabilistic models}, volume = {162}, year = {2025}, }
• (2025). Selecting after sales provider of complex product based on game and matching framework. EAAI, 162, 112524. (www)

  @article{ , abstract = {As a strategic enabler of high-end manufacturing, the high-quality evolution of complex equipment is indispensable for any nation aspiring to industrial leadership. After sales service (AS) long relegated to a support function, which has emerged as a decisive determinant of product life-cycle value and, consequently, of this transformative journey. This study therefore investigates the technological innovation of AS for complex products through a Stackelberg game that captures the collaborative dynamics between an original equipment manufacturer (OEM) and an after-sales service provider (ASP). We derive the necessary and sufficient conditions under which an ASP finds participation economically viable, then embed these conditions into a multi-criteria matching framework that links ASP capabilities with spare-part requirements. Leveraging an entropy weighted DEMATEL (Decision-making Trial and Evaluation Laboratory) hybrid and we first quantify the causal salience of matching attributes and build a parsimonious evaluation index system. Next, by explicitly encoding bilateral attribute preferences, we formulate a two-sided matching model that identifies the Pareto-optimal ASP portfolio for any given product architecture. Finally, backward induction over the integrated game-matching structure yields a prescriptive tool that not only screens ASPs but also prescribes contractual levers to sustain long-term co-innovation. The proposed framework thus unifies strategic participation incentives with operational compatibility, offering OEMs a rigorous, implementable roadmap for selecting and governing after-sales partners in the era of servitized, high-stakes manufacturing.}, archive = {J_EAAI}, author = {Xin Huang and Xiaoyan Qi and Xiaojuan Xu}, doi = {10.1016/j.engappai.2025.112524}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112524}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Selecting after sales provider of complex product based on game and matching framework}, volume = {162}, year = {2025}, }
• (2025). PLDs-CNN-ridge-ELM: Interpretable lightweight waste classification framework. EAAI, 162, 112522. (www)

  @article{ , abstract = {The accelerating global population growth and expanding economic activities have resulted in a notable increase in waste generation, necessitating accurate and efficient waste classification systems for sustainable waste management. This research presents a novel two-stage waste classification model leveraging a Lightweight Parallel Depth-wise Separable Convolutional Neural Network (PLDs-CNN), combined with a Ridge Regression Extreme Learning Machine (Ridge-ELM) classifier, using waste images as input. The proposed system efficiently classifies waste into four primary categories (hazardous, household, recyclable, and residual) in the first stage and further refines the classification into twelve subcategories in the second stage. Featuring a lightweight architecture of nine layers and about 1.09 million parameters, the PLDs-CNN model achieves high accuracy with substantially reduced computational overhead, outperforming many deeper networks. In the four-class classification stage, the system achieves an average accuracy of 99 %, with precision, recall, F1-score, and receiver operating characteristics (ROC)-area under the curve (AUC) values of 97.25 ± 0.02 %, 96 ± 0.03 %, 96.5 ± 0.01 %, and 99.28 %, respectively. In the twelve-class classification, the model continues to deliver superior results, with 96 % accuracy and equally strong precision, recall, and F1-score metrics. The system is supported by a real-time hardware architecture, featuring a user-centric Graphical User Interface (GUI), a webcam-enabled conveyor belt sorting mechanism, and a 2-axis pan-tilt system for automated waste sorting. Additionally, the model's interpretability is significantly improved through the integration of Shapley Additive Explanations (SHAP), which provides important perspectives into the decision-making process, increasing transparency and trustworthiness in real-world applications. The proposed framework not only surpasses conventional methods in both accuracy and computational efficiency but also emphasizes sustainability by facilitating cost-effective and scalable waste management solutions aimed at promoting recycling and resource reuse.}, archive = {J_EAAI}, author = {Mansura Naznine and Md. Nahiduzzaman and Md. Jawadul Karim and Md. Faysal Ahamed and Abdus Salam and Mohamed Arselene Ayari and Amith Khandakar and Azad Ashraf and Mominul Ahsan and Julfikar Haider}, doi = {10.1016/j.engappai.2025.112522}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112522}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {PLDs-CNN-ridge-ELM: Interpretable lightweight waste classification framework}, volume = {162}, year = {2025}, }
• (2025). Interpretable interval prediction of dam displacement based on variational autoencoder and improved temporal fusion transformer considering solar radiation effects. EAAI, 162, 112520. (www)

  @article{ , abstract = {Ensuring the safety of dams is critical to maintaining national economic development and social stability, requiring the implementation of accurate displacement prediction methods for early detection of structural anomalies and effective risk mitigation. However, existing statistical models primarily focus on point predictions, failing to quantify the uncertainty in displacement variations, and often neglect the critical environmental factor of solar radiation. To address these limitations, this study proposes a novel interpretable interval prediction framework that integrates solar radiation factors into an advanced hydrostatic-temperature-time (AHTT) model. A variational autoencoder (VAE) is employed to extract robust latent features from a large volume of measured temperature data, effectively reducing temperature-related noise. Subsequently, an improved temporal fusion transformer method is introduced to probabilistic dam displacement prediction. This method uses an enhanced quantile loss function based on the Huber loss to generate both point and interval predictions that dynamically reflect the prediction uncertainty. In addition, an interpretable multi-head attention module is incorporated to quantify the contribution of each environmental factor. Hyperparameter tuning of the improved temporal fusion transformer is further optimized using Bayesian optimization based on the tree-structured Parzen estimator (TPE), which improves prediction accuracy. Engineering case studies validate that the proposed model not only achieves the highest point prediction accuracy, but also provides narrower prediction intervals with the best coverage width criterion. Ablation experiments and interpretability analyses further confirm the significant impact of solar radiation on dam displacement, providing valuable insights for the development of dam displacement prediction models and risk-informed decision making.}, archive = {J_EAAI}, author = {Taiqi Lu and Hao Gu and Chongshi Gu and Chenfei Shao and Yiming Wang and Dongyang Yuan}, doi = {10.1016/j.engappai.2025.112520}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112520}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Interpretable interval prediction of dam displacement based on variational autoencoder and improved temporal fusion transformer considering solar radiation effects}, volume = {162}, year = {2025}, }
• (2025). An undersampling method for software defect prediction based on hilbert curve mapping distance. EAAI, 162, 112519. (www)

  @article{ , abstract = {The class imbalance problem presents a significant challenge in software defect prediction. The undersampling method enhances prediction performance by eliminating non-defective instances, thereby enabling the model to focus more on defective instances. However, the effective selection of representative non-defective instances while preserving the overall data distribution remains a critical challenge. Inspired by the space-filling property of Hilbert curves, we propose the H ilbert C urve M apping D istance U ndersampling (HCMDU) method for software defect prediction. This method first maps instances to Hamming space to ensure that similar instances are positioned closer together in the space. Instance circular domains are then partitioned based on the Hamming distance between them, which facilitates the exploration of instance variability within a localized region. Finally, the Hilbert curve mapping distance is employed to further uncover the data distribution pattern within the instance circular domains. The experimental results demonstrate that HCMDU delivers outstanding performance across 16 randomly selected software defect datasets in both Random Forest (RF) and Classification and Regression Trees (CART). Moreover, the results are further corroborated by the Friedman ranking and Nemenyi post-hoc test, which indicate that HCMDU significantly improves the performance of software defect prediction.}, archive = {J_EAAI}, author = {Yu Tang and Ye Du and Ang Li and Ming-song Yang and Yan Xia}, doi = {10.1016/j.engappai.2025.112519}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112519}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An undersampling method for software defect prediction based on hilbert curve mapping distance}, volume = {162}, year = {2025}, }
• (2025). Unsupervised fault diagnosis method for rolling bearings based on federated universal domain adaptation. EAAI, 162, 112518. (www)

  @article{ , abstract = {To address the issues of low diagnostic model accuracy caused by non-sharing of rolling bearing private data, distribution differences, and label space discrepancies across multiple clients, as well as the challenges that certain clients face in obtaining labeled data, an unsupervised fault diagnosis method is proposed for rolling bearings based on federated universal domain adaptation (FUDA). First, privacy protection during the transmission process in federated learning is ensured by implementing random mapping at local clients. Second, the central server employs the proposed mixed radial basis kernel-maximum mean discrepancy (MR-MMD) method to further mitigate distributional disparities between the feature spaces of source and target clients. This achieves unsupervised features alignment between these features. Third, margin vectors are introduced to tackle label space disparities between source and target clients, enabling effective separation of unknown class samples in the dataset of the target client. Finally, a dynamic weighted loss fusion strategy is designed to adaptively optimize the weight ratios of different losses. This enhancement facilitates the learning efficiency of the model. Experimental validation on two datasets demonstrates that the proposed approach can achieve average accuracies of 95.6 % and 87.7 % for the respective datasets. Compared with other methods, it represents improvements of 6.5 % and 8.1 %, while training time is reduced by at least 27 %. These results validate the effectiveness of the proposed method.}, archive = {J_EAAI}, author = {Shouqiang Kang and Yulin Sun and Xinrui Li and Yujing Wang and Qingyan Wang and Xintao Liang}, doi = {10.1016/j.engappai.2025.112518}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112518}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unsupervised fault diagnosis method for rolling bearings based on federated universal domain adaptation}, volume = {162}, year = {2025}, }
• (2025). A multi-scale feature extraction and fusion framework based on wavelet Kolmogorov–Arnold networks and parallel bi-directional gated recurrent units for electric load forecasting. EAAI, 162, 112517. (www)

  @article{ , abstract = {Short-term electric load forecasting remains challenged by the dual requirements of accuracy and robustness due to the combined effects of strong seasonality, multi-scale spikes, and stochastic disturbances. To address this, we propose a novel multi-scale forecasting framework, NP-WavKAN-Fusion, which integrates Neural Prophet for data decomposition and a Wavelet-based Kolmogorov–Arnold Network (WavKAN) with learnable wavelet kernels for multi-scale encoding. This fusion model utilizes a Bi-directional Gated Recurrent Unit (BiGRU) to capture long-term temporal dependencies and an adaptive feature fusion gate (AFF) to dynamically re-weight static and dynamic features for final load predictions. Extensive experiments on two public datasets from Australia and Morocco show that NP-WavKAN-Fusion consistently outperforms traditional models, reducing the mean absolute error by at least 30 %. For multi-step forecasting tasks, NP-WavKAN-Fusion maintains error inflation within 15 %, demonstrating superior performance compared to state-of-the-art long-sequence models such as Informer and PatchTST. The Diebold–Mariano test confirms that NP-WavKAN-Fusion yields statistically significant improvements, with 19 out of 20 comparisons showing lower errors. Ablation studies show that removing either the Neural Prophet component or the AFF significantly increases the forecasting error, validating the necessity of our layered denoising and fusion strategies. The proposed NP-WavKAN-Fusion framework demonstrates strong potential for real-world applications in electric load forecasting, offering robust performance under various temporal and non-stationary conditions.}, archive = {J_EAAI}, author = {Chunliang Mai and Lixin Zhang and Xuewei Chao and Xue Hu and Omar Behar}, doi = {10.1016/j.engappai.2025.112517}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112517}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-scale feature extraction and fusion framework based on wavelet Kolmogorov–Arnold networks and parallel bi-directional gated recurrent units for electric load forecasting}, volume = {162}, year = {2025}, }
• (2025). Resilient kernel-based unsupervised multi-view feature selection via compact binary hashing. EAAI, 162, 112515. (www)

  @article{ , abstract = {Multi-view feature selection across diverse views identifying a compact subset of the most informative feature across various data views without relying on labeled information. While most of the solutions are limited to linear multi-view data or utilize weakly-supervised single-label learning to assist in feature selection, leading to the loss of valuable semantic information, especially when dealing with complex real-world multi-view datasets. To overcome these limitations, we introduce a novel Resilient Kernel-based Unsupervised Multi-view Feature Selection via compact Binary Hashing (RKUMBH), which aims to search a robust and consistent graph representation across views, leveraging binary hashing codes to guide feature selection. Specifically, we first standardize the dimensionality of multi-view data by using non-linear kernel mapping. Then, we explore consistent graph structures across different views by fusing individual similarity graph of each view under a self-representation guidance. Moreover, the low-rank constraints are used to preserve the primary structures and patterns embedding within the data, and an unsupervised hashing feature selection framework is conducted to generate reliable hashing codes across views. Additionally, we design a customized iterative optimization method to solve the unified model. Extensive experiments on six public multi-view datasets demonstrate that our proposed method obtains state-of-the-art results compared to existing works for both clustering and feature selection tasks.}, archive = {J_EAAI}, author = {Rongyao Hu and Mengmeng Zhan and Jiangzhang Gan and Li Li and Fei Ye and Tong Liu}, doi = {10.1016/j.engappai.2025.112515}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112515}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Resilient kernel-based unsupervised multi-view feature selection via compact binary hashing}, volume = {162}, year = {2025}, }
• (2025). Uncovering the geometry-dependent optical asymmetry of gold nanorods helical assemblies using artificial neural networks. EAAI, 162, 112513. (www)

  @article{ , abstract = {The optical asymmetry of gold nanorods (Au-NRs) helical assemblies is well-documented with a wide range of applications. Nevertheless, the geometry-dependent optical asymmetry within these assemblies has not been adequately explored and quantified. The present study proposes a novel approach to predict the optical asymmetry of Au-NRs helical assemblies based on geometric characteristics using artificial neural networks (ANN). The performance of the ANN termed 3 N H L 50 N N was significantly enhanced through the optimization of the hidden layer and node, resulting in an R 2 of the outcomes exceeding 0.998 and a reduction in computational time exceeding 99.99 %. In instances where the specific geometric characteristics are needed to attain a desired optical asymmetry, a retrieval of geometric characteristics of Au-NRs helical assemblies was additionally investigated using a traversing mechanism featured particle swarm optimization (PSO) algorithm. The results of the retrieval were obtained within 6 s and demonstrate a high degree of accuracy and reliability. The combination of the 3 N H L 50 N N and the PSO algorithm is capable of accurately predicting the optical asymmetry of Au-NRs helical assemblies and the retrieval of the geometry characteristics, thereby enabling the quantitative understanding of their overall geometry-dependent optical asymmetry.}, archive = {J_EAAI}, author = {Yang Liu and Yongguang Chen and Xiyang Wei and Jianhua Shang and Lina Zhao}, doi = {10.1016/j.engappai.2025.112513}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112513}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Uncovering the geometry-dependent optical asymmetry of gold nanorods helical assemblies using artificial neural networks}, volume = {162}, year = {2025}, }
• (2025). Detecting worker loss of balance events from point cloud sequence using unsupervised motion-pose learning. EAAI, 162, 112512. (www)

  @article{ , abstract = {Workers' loss of balance (LB), such as slip and trip, may lead to severe injuries and even fatalities. Existing methods for detecting LB typically rely on wearable sensors and focus on specific body parts. This study introduces a novel, non-contact approach utilizing light detection and ranging (LiDAR) technology to detect LB events. By capturing full-body point cloud data, the proposed method extracts both static pose and dynamic motion features across multiple body sections and detects LB events through unsupervised learning. The high-dimensional point cloud sequence is transformed into interpretable gait features, enabling effective unsupervised learning through sequence reconstruction. A two-stream network and fusion strategy are also developed to combine pose and motion features for final LB detection. Experiments with various LB events demonstrate the method's effectiveness, achieving an F1 score of 0.98 and a recall of 0.98. Our analysis reveals that integrating features from multiple body parts and the fusion of pose and motion information significantly enhances detection performance. This study offers a promising alternative to traditional methods, providing effective, non-intrusive monitoring of worker safety in dynamic construction environments.}, archive = {J_EAAI}, author = {Mingyu Zhang and Lei Wang and Yinong Hu and Shuai Han and Jiawen Zhang and Heng Li}, doi = {10.1016/j.engappai.2025.112512}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112512}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Detecting worker loss of balance events from point cloud sequence using unsupervised motion-pose learning}, volume = {162}, year = {2025}, }
• (2025). Human motion prediction using mixture-of-branch graph convolutional network. EAAI, 162, 112511. (www)

  @article{ , abstract = {Using the same spatio-temporal feature extraction network to predict multiple types of human motions pose a challenge for the prediction model to achieve optimal performance. To address this issue and achieve differentiated training, we propose a novel Mixture-of-Branch Graph Convolutional Network model which simultaneously inserts human motion sequences into a multi-branch human motion prediction module and a branch weight allocation module. When generating the final prediction sequence, weights are assigned to the prediction results of each branch. Mixture-of-Branch Graph Convolutional Network employs loss values to control competition rather than cooperation among branches, effectively addressing the issue of mutual influence between sequences during network training. To the best of our knowledge, this marks the inaugural utilization of a Mixture-of-Branch network in the realm of human motion prediction. To optimize the efficiency of the multi-branch model and reduce prediction complexity, we introduce a spatio-temporal feature extraction method for the human skeleton that accommodates Euclidean geometric transformations. This method liberates the Mixture-of-Branch Graph Convolutional Network from the constraints of additional branches, allowing it to handle similar motion sequences under varying degrees of translation or rotation, where feature matrices may exhibit significant differences. The proposal of Mixture-of-Branch Graph Convolutional Network and its related experiments represent our contribution to the Artificial Intelligence field, with significant potential value in engineering applications as well. Mixture-of-Branch Graph Convolutional Network is tested on the Human3.6M, Carnegie Mellon University Motion Capture, and Three-Dimensional Human Pose in the Wild datasets, achieving high performance. Particularly noteworthy is the 7% overall performance improvement in Mean Per Joints Position Error prediction on the Carnegie Mellon University Motion Capture dataset.}, archive = {J_EAAI}, author = {Xianshan Li and Ang Gao and Xingxing Ning and Fengda Zhao}, doi = {10.1016/j.engappai.2025.112511}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112511}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Human motion prediction using mixture-of-branch graph convolutional network}, volume = {162}, year = {2025}, }
• (2025). A hybrid multivariate normal boundary intersection approach with post-optimization assisted by mixture design of experiments. EAAI, 162, 112510. (www)

  @article{ , abstract = {This article presents a hybrid multiobjective optimization approach for identifying Pareto-optimal solutions by combining evaluation metrics. Using response surface models built from rotated factor scores, the algorithm analyzes the influence of weights on solutions through performance metrics, applying a weighted combination as a post-optimization strategy. The methodology introduces innovations such as the combination of independent functions, polynomial analysis to discover intermediate weights, and the use of prediction-based metrics. To illustrate application of the methodology, a case study was conducted on the turning process of tempered cylindrical steel bars, evaluating dimensions of reliability, quality, and economic performance. The best solution obtained included the following input variables: a cutting speed of 147.792 m per minute, a feed rate of 0.156 mm per revolution, and a depth of cut of 0.247 mm. The resulting output variables included a mean tool life of 46.316 min, mean time to failure of 47.164 min, a wear rate of 0.006 mm per minute, an average surface roughness of 0.746 μm, a total surface roughness of 3.406 μm, a process cost of 3.721 dollars, a return on investment of 0.187, and overall equipment effectiveness of 0.387. Comparisons were made with other optimization methods, revealing satisfactory performance based on key metrics. The methodology was also tested on benchmark functions to assess its robustness and adaptability to different scenarios. In addition, the method was implemented in an online quality monitoring system using the Random Forest machine learning algorithm, and results indicate that optimal conditions can be identified quickly and effectively.}, archive = {J_EAAI}, author = {Matheus Costa Pereira and Caio Tertuliano Ribeiro and Ronã Rinston Amaury Mendes and Paulo Henrique da Silva Campos and Anderson Paulo de Paiva}, doi = {10.1016/j.engappai.2025.112510}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112510}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A hybrid multivariate normal boundary intersection approach with post-optimization assisted by mixture design of experiments}, volume = {162}, year = {2025}, }
• (2025). Inverse design of particle shapes with target sphericity and packing fraction using variational autoencoders. EAAI, 162, 112509. (www)

  @article{ , abstract = {Sphericity and packing fraction are fundamental properties governing the behavior of granular materials in many engineering applications. Conventional methods for designing particles with these target properties usually suffer from limited accuracy, diversity, and interpretability due to complex relationships between particle shape and properties. To address this, we propose an inverse design framework based on deep learning. First, a rotation- and reflection-invariant variational autoencoder (VAE) parameterizes two-dimensional convex particle shapes into a low-dimensional latent space, enabling accurate reconstruction and capturing geometric interpretations such as sphericity and symmetry. Second, a conditional variational autoencoder (CVAE) facilitates inverse design by generating particle shapes corresponding to target sphericity or packing fraction, and also enables the coupling control of both properties. Trained on a dataset of over 1600 convex shapes, the framework demonstrates robustness and universality. The rotation- and reflection-invariant architecture consistently maps different orientations of the same shape to a unified representation, which enhances interpretability. The main contribution in artificial intelligence lies in developing invariant generative models that learn shape representations and enable property-driven shape generation. The engineering contribution is providing a precise and efficient tool for the inverse design of particle shapes with target properties, supporting the optimization of granular materials in engineering applications.}, archive = {J_EAAI}, author = {Yutong Qian and Shuixiang Li}, doi = {10.1016/j.engappai.2025.112509}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112509}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Inverse design of particle shapes with target sphericity and packing fraction using variational autoencoders}, volume = {162}, year = {2025}, }
• (2025). Three-way dynamic clustering algorithms based on generalized neighborhood relations in incomplete hybrid information systems with applications in medical decision-making. EAAI, 162, 112508. (www)

  @article{ , abstract = {In addressing clinical challenges related to chronic diseases, the effective use of medical data in decision-making is often hindered by issues such as incompleteness, heterogeneity, and the need for continuous updates. To cope with these challenges, this study introduces a three-way dynamic clustering strategy built upon generalized neighborhood relations, aiming to enhance clustering robustness, strengthen the model’s ability to manage uncertainty, and support adaptability to dynamically evolving data. First, generalized neighborhood relations are constructed in incomplete hybrid information systems. An evaluation function is defined from two perspectives: the number of similar attributes between objects and the distance between objects, thereby optimizing similarity measurement and accurately characterizing the data structure. Second, three-way decision rules are introduced to effectively handle uncertainty in objects while maintaining classification accuracy, thereby improving the interpretability and adaptability of the clustering model. Furthermore, to accommodate the dynamic nature of medical data, a dynamic incremental clustering method based on neighborhood information is proposed to ensure that newly added patient data can be efficiently integrated into existing clusters, enhancing model real-time performance and computational efficiency. Experiments conducted on real clinical data from Chronic kidney disease (CKD) patients validate the proposed method. The results demonstrate that, compared to existing clustering algorithms, the proposed method outperforms in terms of F1-score and Rand Index evaluation metrics. It also exhibits higher applicability in patient classification, core and boundary domain partitioning, and dynamic data processing, providing effective support for precision stratified management of chronic disease patients and intelligent medical decision-making.}, archive = {J_EAAI}, author = {Haoran Sun and Bingzhen Sun and Xixuan Zhao and Qiang Bao and Xiaoli Chu}, doi = {10.1016/j.engappai.2025.112508}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112508}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Three-way dynamic clustering algorithms based on generalized neighborhood relations in incomplete hybrid information systems with applications in medical decision-making}, volume = {162}, year = {2025}, }
• (2025). A three-stage segmentation framework for lung cancer lesion isolation in three-dimensional positron emission tomography images. EAAI, 162, 112507. (www)

  @article{ , abstract = {Background Positron emission tomography (PET) is a critical functional medical imaging modality for the early detection and diagnosis of cancers. PET imaging faces several challenges that hinder accurate interpretation including its inherently low spatial resolution, substantial variability in cancer lesions’ appearance, and difficulties distinguishing between the image background and benign lesions. Methods We propose a novel three-stage image segmentation framework to enhance the accuracy of lung cancer lesion identification and extraction from three-dimensional (3D) PET images. The first stage conducts a coarse segmentation using an encoder-decoder structure network to roughly position lesions. The second stage employs a multi-layer feature extraction network to learn the detailed characteristics of coarse segmentation results, mitigating false positives caused by localization inaccuracy. The last stage further refines the extracted features via dividing a sub-region of the lesion into foreground and background branches, reducing false positives caused by over-segmentation of edges. A novel lesion count loss function is introduced to guide the model to generate predictions during the training, ensuring that the predicted lesion counts align with the ground truth labels. Results The proposed method was evaluated on clinical 3D PET image datasets. Experimental results demonstrated a Dice Similarity Coefficient (DSC) of 85.35 %, Accuracy of 83.97 %, and Recall of 86.83 %. Compared to existing models applied to the same datasets, our method consistently achieved superior performance. Conclusion The proposed method significantly improves the segmentation performance of lung cancer lesions, implying that our method holds substantial potential for broader clinical application, even in low-resolution images.}, archive = {J_EAAI}, author = {Yusheng Wu and Qiang Lin and Jingjun Wei and Yongchun Cao and Zhengxing Man and Xiaodi Huang}, doi = {10.1016/j.engappai.2025.112507}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112507}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A three-stage segmentation framework for lung cancer lesion isolation in three-dimensional positron emission tomography images}, volume = {162}, year = {2025}, }
• (2025). Self-supervised visual assessment of railway track curvature via homography learning based on projective curvilinear geometry model. EAAI, 162, 112506. (www)

  @article{ , abstract = {Railway track curvature monitoring is crucial for ensuring operational safety and passenger comfort. As a robust complement to the single-point physical sensor approaches, vision-based methods have recently gained increasing adoption. However, existing approaches frequently neglect the systematic exploitation of railway ego-vision geometry in two critical aspects: (1) the rail-camera kinematic coupling that relates the rail appearance in camera’s view and the track curvature during curvilinear motion, and (2) the potential of self-supervised learning to overcome annotation scarcity in this domain. This geometric oversight limits their accuracy in real-world dynamic scenarios. To address these gaps, this study proposes a novel vision-based framework that systematically exploits the railway ego vision geometry. Our methodology comprises two key innovations: First, a projective curvilinear geometry model that mathematically relates the ground-planes-induced homography to actual track curvature, thereby establishing a mapping from curvature and its variation to rail imaging curves. Second, a self-supervised curvature prediction network trained using automatically generated labels from our geometric model, eliminating the need for manual curve annotations. The self-supervision is achieved through a cyclic consistency mechanism between predicted curvatures and reprojected image features. Experimental validation using real-world railway footage demonstrates significant improvements: Our method reduces the average root mean squared error by 23.31% compared to state-of-the-art vision-based curvature estimation methods. These results underscore the effectiveness of geometry-aware computer vision for railway geometry monitoring}, archive = {J_EAAI}, author = {Peng Tang and Zhibin Yu}, doi = {10.1016/j.engappai.2025.112506}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112506}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Self-supervised visual assessment of railway track curvature via homography learning based on projective curvilinear geometry model}, volume = {162}, year = {2025}, }
• (2025). Saliency and correlation learning for co-salient object detection. EAAI, 162, 112504. (www)

  @article{ , abstract = {Co-Salient object detection aims to identify common salient objects across a given group of images. However, accurately locating co-salient objects remains challenging due to the complexity of capturing the correlation representation of each group of images. To tackle this problem, we propose a saliency and correlation learning method for co-salient object detection. This method employs a saliency learning network and a correlation learning network to generate precise co-saliency maps of a group of images. Within the saliency learning network, a saliency feature grafting module is designed to refine object edges and achieve accurate detection of salient objects. Furthermore, the correlation learning network incorporates two modules, which are designed for extracting saliency correlation representation and deriving consensus correlation representation within a group of images, respectively. Guided by prior information obtained from saliency learning of images, our method significantly improves performance in co-salient object detection through correlation representation learning. Extensive experiments on all the latest benchmarks demonstrate that our method outperforms 11 state-of-the-art models, achieving a new level of technical excellence, with an average Structural Similarity Measure score of 0.845.}, archive = {J_EAAI}, author = {Ying Tong and Xiangfeng Luo and Liyan Ma and Shaorong Xie}, doi = {10.1016/j.engappai.2025.112504}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112504}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Saliency and correlation learning for co-salient object detection}, volume = {162}, year = {2025}, }
• (2025). Automated monitoring of mycobacterium population growth in time-lapse microscopy with deep learning pseudo-supervised algorithm. EAAI, 162, 112503. (www)

  @article{ , abstract = {The understanding of bacteria evolution through time and the analysis of their interaction with the medium is a main step in biomedical research. Bacteria cause in humans different infectious diseases, such as tuberculosis. Tuberculosis is one of the leading causes of death worldwide, caused by Mycobacterium tuberculosis ( Mtb ). Observation of Mycobacteria under a microscope provides essential information for disease-fighting. Nevertheless, Mtb presents a very particular cell division pattern characterized by a high entanglement in which the frontier between dividing cells are almost indistinguishable to the human eye, which makes the manual analysis of cultures a very challenging task, resulting in extremely long processing times. In this paper, we propose a novel approach to automatically detect bacterial cells and infer the growth rate and growth speed over a time-lapse microscopy (TLM) sequence. This method leverages a semi-supervised Deep Learning approach with a lightweight U-Net for efficient and fast processing. To reduce efforts of database construction, we deploy and exploit a mix of manually and synthetically annotated databases of two different classes of Mycobacterium , i.e. Mtb and Mycobacterium smegmatis ( Msm ). The experiments demonstrated the efficiency of the detection of both bacterial reproduction rate and bacterial death in Mycobacterium species, with the Phase-contrast (PhC) microscopy channels as a unique input. Reaching a segmentation accuracy of 96.3%, recall of 78.2%, and precision of 85.5% on Mtb detection for the model trained with a mix of manually and synthetically annotated databases. The experiments validate as well the synthetic label generation technique, which provides more conscience and accurate results than the manually segmented techniques. The mask generation workflow opens new opportunities for the microbial AI development field. The system implemented may change the paradigm in preclinical anti-mycobacteria drug development for since enables the adoption of TLM, a technology that provides extremely valuable information on drug efficacy but is seldom leveraged due to the complex image processing it involves.}, archive = {J_EAAI}, author = {Lara Visuña and Javier Garcia-Blas and Santiago Ferrer-Bazaga and Patricio Lopez-Exposito and Jesus Carretero}, doi = {10.1016/j.engappai.2025.112503}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112503}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Automated monitoring of mycobacterium population growth in time-lapse microscopy with deep learning pseudo-supervised algorithm}, volume = {162}, year = {2025}, }
• (2025). Reducing the estimation bias and variance in reinforcement learning via maxmean and aitken value iteration. EAAI, 162, 112502. (www)

  @article{ , abstract = {The value-based reinforcement leaning methods suffer from overestimation bias, because of the existence of max operator, resulting in suboptimal policies. Meanwhile, variance in value estimation will cause the instability of networks. Many algorithms have been presented to solve the mentioned, but these lack the theoretical analysis about the degree of estimation bias, and the trade-off between the estimation bias and variance. Motivated by the above, in this paper, we propose a novel method based on Maxmean and Aitken value iteration, named MMAVI. The Maxmean operation allows the average of multiple state–action values (Q values) to be used as the estimated target value to mitigate the bias and variance. The Aitken value iteration is used to update Q values and improve the convergence rate. Based on the proposed method, combined with Q-learning and deep Q-network, we design two novel algorithms to adapt to different environments. To understand the effect of MMAVI, we analyze it both theoretically and empirically. In theory, we derive the closed-form expressions of reducing bias and variance, and prove that the convergence rate of our proposed method is faster than the traditional methods with Bellman equation. In addition, the convergence of our algorithms is proved in a tabular setting. Finally, we demonstrate that our proposed algorithms outperform the state-of-the-art algorithms in several environments.}, archive = {J_EAAI}, author = {Fanghui Huang and Wenqi Han and Xiang Li and Xinyang Deng and Wen Jiang}, doi = {10.1016/j.engappai.2025.112502}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112502}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Reducing the estimation bias and variance in reinforcement learning via maxmean and aitken value iteration}, volume = {162}, year = {2025}, }
• (2025). Hybrid reinforcement learning in parameterized action space via fluctuates constraint. EAAI, 162, 112499. (www)

  @article{ , abstract = {Parameterized actions in Reinforcement Learning (RL) are composed of discrete-continuous hybrid action parameters, which are widely employed in game scenarios. However, previous works have often concentrated on the network structure of RL algorithms to solve hybrid actions, neglecting the impact of fluctuations in action parameters for agent move trajectory. Due to the coupling between discrete and continuous actions, instability in discrete actions influences the selection of corresponding continuous parameters, resulting in the agent deviating from the optimal move path. In this paper, we propose a parameterized RL approach based on parameter fluctuation restriction (PFR) to address this problem, called CP-DQN. Our method effectively mitigated value fluctuation in action parameters by constraining the action parameter between adjacent time steps. Additionally, we have incorporated a supervision module to optimize the entire training process. To quantify the superiority of our approach in minimizing trajectory deviations for agents, we propose an indicator to measure the influence of parameter fluctuations on performance in hybrid action space. Our method is evaluated in three environments with hybrid action spaces, and the experiments demonstrate the superiority of our method compared to existing approaches.}, archive = {J_EAAI}, author = {Chengcheng Yan and Shujie Chen and Jiawei Xu and Xuejie Wang and Zheng Peng}, doi = {10.1016/j.engappai.2025.112499}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112499}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hybrid reinforcement learning in parameterized action space via fluctuates constraint}, volume = {162}, year = {2025}, }
• (2025). MAVREN: A multilayered adaptive framework for deploying vision-language models on resource-constrained unmanned aerial vehicles for autonomous search and rescue. EAAI, 162, 112498. (www)

  @article{ , abstract = {Unmanned Aerial Vehicles (UAVs) have become indispensable in autonomous search and rescue (SAR) missions, where the ability to interpret complex visual scenes in real time is critical. When equipped with artificial intelligence (AI)-empowered Vision-Language Models (VLMs), UAVs can provide rich contextual insights, interpret their findings, and even suggest next steps, but their deployment on resource-constrained UAV platforms is vastly limited by high computational demands, energy constraints, and strict latency requirements. This paper introduces MAVREN, a m ultilayered a daptive scheduler for V LM execution in re source-constrained UAV n etworks for autonomous SAR operations. Evaluations conducted on NVIDIA Jetson Orin NX using state-of-the-art VLMs such as Large Language and Vision Assistant (LLaVA) 1.6 and Vision-Language Alignment (VILA) 7B demonstrate that MAVREN achieves up to 26.11% higher throughput , 23% lower energy consumption , 13.51% reduced latency , and a 7% gain in detection accuracy compared to baseline schedulers across indoor, outdoor, and multi-UAV SAR scenarios. This is achieved through the integration of a visual encoder for lightweight feature extraction, a block floating-point quantizer for precision-efficient representation, a bit-wise computation engine for fast arithmetic execution, and a branch-and-bound optimizer for dynamic central processing unit (CPU) scheduling. These tightly coupled components allow MAVREN to optimize the energy–latency–accuracy trade-off, making it a deployable solution for vision-language reasoning in real-world SAR missions. Our findings demonstrate MAVREN’s capability to deliver rapid, energy-efficient inference, advancing the deployment of computationally intensive VLMs on resource-constrained UAV platforms.}, archive = {J_EAAI}, author = {Md Tahmid Rashid and Md Jawad Siddique and Abdus Shaqur}, doi = {10.1016/j.engappai.2025.112498}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112498}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {MAVREN: A multilayered adaptive framework for deploying vision-language models on resource-constrained unmanned aerial vehicles for autonomous search and rescue}, volume = {162}, year = {2025}, }
• (2025). Rapid prediction of structural deflection based on explainable machine learning. EAAI, 162, 112497. (www)

  @article{ , abstract = {High-pressure arched air-rib membrane structures (HP-ARMS) exhibit lightweight portability and modular installation. This study proposes an explainable machine learning (ML) framework for the prediction of blast-induced HP-ARMS deflection. Firstly, compare the simulation results with the experimental data to verify the numerical method accuracy. Subsequently, a database containing 500 samples was established through numerical modeling. The input features include 6 structural parameters (air-rib pressure (X1), air-rib diameter (X2), air-rib thickness (X3), air-rib width (X4), air-rib height (X5), and air-rib bottom consolidation method (X6)) and 4 external load parameters (soil cover depth (X7), lateral explosion distance (X8), explosion equivalent (X9), and charge burial depth (X10)). Use Six ML algorithms—Light Gradient Boosting Machine (LightGBM), Random Forest (RF), Adaptive Boosting (AdaBoost), K-Nearest Neighbors (KNN), Convolutional Neural Network (CNN), and Gradient Boosting (GB)— and use four evaluation metrics to assess the accuracy of the ML model. Finally, the SHapley Additive exPlans (SHAP) method was used for interpretable analysis. The results showed that the LightGBM model had the best prediction performance. Compared with LightGBM, the Random-LightGBM (R-LightGBM) model significantly improved performance after hyperparameter optimization, with Root Mean Square Error (RMSE) reduced by 2.75 %, Mean Absolute Percentage Error (MAPE) reduced by 8.16 %, Mean Absolute Error (MAE) reduced by 5.88 %, and R-Squared (R 2 ) increased by 0.01. The SHAP method indicates that the explosion equivalent (X9) and charge burial depth (X10) are the most important parameters.}, archive = {J_EAAI}, author = {Yongtao Mi and Yushuai Zhang and Yicun Chen and Chenxi Sun and Huiqi Ren}, doi = {10.1016/j.engappai.2025.112497}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112497}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Rapid prediction of structural deflection based on explainable machine learning}, volume = {162}, year = {2025}, }
• (2025). A latent-coupled neural network for multiphysics long-term forecasting in reactor transients using sparse observations. EAAI, 162, 112496. (www)

  @article{ , abstract = {Complex dynamical systems in safety-critical applications like nuclear reactors involve strongly coupled physical fields evolving over space and time. Accurate prediction of these fields is vital for safety monitoring but is challenged by limited sensor placement and unobservable variables ( e.g. , xenon and iodine concentrations). This paper proposes the S parse observation to H igh-dimensional coupled physical field P rediction Network (SHPNet), a deep learning framework that predicts and reconstructs multiple physical fields directly from sparse observations.SHPNet combines a three-branch autoencoder to extract shared latent representations with a neural operator that models temporal dynamics in latent space, enabling efficient long-term forecasting. Evaluated on H ua-long P ressurized R eactor (HPR1000) under varying power and burnup conditions, SHPNet outperforms traditional frameworks and end-to-end model , achieving higher accuracy, robustness to observation sparsity, and effective reconstruction of unobservable fields. These results demonstrate SHPNet’s potential as a practical tool for real-time monitoring of complex coupled systems.}, archive = {J_EAAI}, author = {Yu-Yan Xu and Jun Luo and Deng Pan and Wei Lu and Ting Liu and Guanghui Yuan and Minxiao Zhong and Qing Li and Helin Gong}, doi = {10.1016/j.engappai.2025.112496}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112496}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A latent-coupled neural network for multiphysics long-term forecasting in reactor transients using sparse observations}, volume = {162}, year = {2025}, }
• (2025). ESIGCF: Extremely simplified but intent-enhanced graph collaborative filtering for recommendation. EAAI, 162, 112495. (www)

  @article{ , abstract = {Graph convolutional networks (GCNs) and graph contrastive learning (GCL) have substantially advanced recommender systems by modeling high-order user–item interactions and leveraging self-supervised signals. However, many existing methods overemphasize user–user or item–item similarities and rely on complex intent modeling, leading to increased complexity and limited exposure to diverse items. To address these challenges, we propose ESIGCF ( E xtremely S implified but I ntent-enhanced G raph C ollaborative F iltering) — a lightweight yet effective recommendation framework. ESIGCF explicitly defines user intent as the inner product between user and item embedding vectors and comprises two primary modules: (i) an intent-enhanced GCN that uses hybrid normalization (combining mean- and symmetric-normalization) to capture fine-grained user–item preferences without additional intent parameters, and (ii) an intent-aware GCL that aligns user–item pairs and positive and generated negative items. Negative samples are generated via a non-linear activation of item embedding interactions, promoting exposure to varied candidates without data augmentation. Experiments on three public datasets (Alibaba-iFashion, Yelp2018, Amazon-Book) show that ESIGCF consistently outperforms state-of-the-art baselines. For instance, on Alibaba-iFashion, ESIGCF achieves Recall@20 of 0.1273 versus 0.1059 for the best intent-enhanced baseline (a 20.2% relative improvement). Comprehensive experiments confirm that ESIGCF effectively captures latent user intent, mitigates popularity bias, and enhances recommendation performance with reduced complexity. Our code is available at https://github.com/Yangzhi22/ESIGCF .}, archive = {J_EAAI}, author = {Zhi Yang and Ruizhang Huang and Yanping Chen and Chuan Lin and Yongbin Qin}, doi = {10.1016/j.engappai.2025.112495}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112495}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {ESIGCF: Extremely simplified but intent-enhanced graph collaborative filtering for recommendation}, volume = {162}, year = {2025}, }
• (2025). Fusing spatial–temporal information into deep learning via wind propagation theory to enhance wind power prediction. EAAI, 162, 112494. (www)

  @article{ , abstract = {Accurately predicting wind power poses significant challenges because of the inherent randomness and intermittency of wind speed, thereby impeding effective wind power scheduling. This study proposes an improved deep learning model which leverages wind propagation theory to uncover spatial–temporal relationships among wind turbines to enhance the performance of wind power prediction. In addition, comprehensive theoretical and empirical analyses are conducted to justify the effectiveness of leveraging wind propagation theory for capturing spatio-temporal relationships among wind turbines. Moreover, spatio-temporal dependencies are modeled through a dual mechanism: multi-channel independent modeling for per-turbine temporal dynamics and wind propagation-based matrix computations for inter-turbine spatial relationships, which together significantly reduce computational complexity while preserving predictive performance. Data from 134 wind turbines and six comparison models were employed to validate the robustness and effectiveness of the proposed model. Empirical results indicate that the proposed model outperforms the baseline models, achieving an average improvement of 6.19% in Root Mean Square Error and 7.05% in Mean Absolute Error.}, archive = {J_EAAI}, author = {Maolin He and Jujie Wang}, doi = {10.1016/j.engappai.2025.112494}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112494}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fusing spatial–temporal information into deep learning via wind propagation theory to enhance wind power prediction}, volume = {162}, year = {2025}, }
• (2025). LIP-MC: Multi-constraint label independent prediction in label distribution learning. EAAI, 162, 112493. (www)

  @article{ , abstract = {Label distribution learning can resolve label ambiguity precisely by determining how well each label describes an instance. Traditional label distribution learning algorithms frequently attempt to model the complex relationships between all labels. This not only adds complexity to the model but may also reduce prediction accuracy due to label conflicts. In this paper, we propose a novel label distribution learning algorithm based on Multi-Constraint Label Independent Prediction (LIP-MC), intended to promote the rationality and accuracy of prediction results by simplifying the prediction process and combining multiple constraints. Specifically, label independent prediction values are generated for each label through sparsity constraints and weight coefficient matrices. Subsequently, a novel transformation model is designed to combine all separate label predictions and produce the final label distribution. Furthermore, smoothness constraints and logarithmic similarity constraints were introduced to enhance the model’s performance and generalization ability. On fourteen real datasets, the experiment was carried out, and the comparison results against seven advanced algorithms under seven evaluation metrics confirmed that the proposed algorithm is superior.}, archive = {J_EAAI}, author = {Gui-Lin Li and Ruili Wu and Xiaorui Qian and Qiang Zhu and Heng-Ru Zhang}, doi = {10.1016/j.engappai.2025.112493}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112493}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {LIP-MC: Multi-constraint label independent prediction in label distribution learning}, volume = {162}, year = {2025}, }
• (2025). Physics-informed dynamic bayesian networks for time-dependent reliability prediction of subsea wellhead sealing system with multi-states. EAAI, 162, 112492. (www)

  @article{ , abstract = {The Subsea Wellhead Sealing System (SWSS) is crucial for the safety of deepwater operating, yet its reliability assessment faces challenges from harsh environments and multi-factor interactions. This study developed a data-driven, physics-informed reliability assessment method combining Finite Element Analysis (FEA) and Dynamic Bayesian Networks (DBN). An FEA model is established based on metal sealing theory, and a data-driven reliability model is subsequently constructed through sampling analysis, with a numerical-to-state conversion method bridging FEA and DBN. The FEA-DBN approach offers two key advantages: eliminating expert scoring subjectivity through physics-based modeling and effectively capturing multi-factor interactions and time-dependent behaviors. Results show this method can precisely quantify the evolution of SWSS reliability throughout its service lifecycle, with the probability of failure increasing from 0.64 % to 3.38 % over a 30-year service life. Case studies demonstrate its effectiveness for deep-sea equipment assessment, particularly in operating environments where real-time monitoring proves challenging, thereby demonstrating significant engineering application value.}, archive = {J_EAAI}, author = {Shengnan Wu and Han Gong and Long Yu and Aibo Zhang and Laibin Zhang and Yiliu Liu}, doi = {10.1016/j.engappai.2025.112492}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112492}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Physics-informed dynamic bayesian networks for time-dependent reliability prediction of subsea wellhead sealing system with multi-states}, volume = {162}, year = {2025}, }
• (2025). Constructing a robust short-text clustering model for contrastive learning based on optimized adaptive optimal transport for pseudo-label generation. EAAI, 162, 112491. (www)

  @article{ , abstract = {Short-text data often suffers from noise and class imbalances, posing challenges for effective clustering. To address these issues, we propose a Short-Text Clustering Model based on Pseudo-Labels and Contrastive Learning (SCPCL). The model comprises two key components: (1) a pseudo-label acquisition module, which introduces the optimal transport theory into short-text clustering and adopts a dynamically adjusted prior distribution to enhance the clustering of minority classes; and (2) a contrastive learning module combining a supervised clustering network, an instance contrastive head, and an anchor network. These components ensure intraclass compactness, interclass separability, and robustness to noise. Experiments on six benchmark datasets showed that SCPCL achieves an average clustering accuracy improvement of 2.61%, with a maximum gain of 6.47% for long-tailed distributions. This model provides an effective solution for clustering complex short text data.}, archive = {J_EAAI}, author = {Jiahui Liu and Chun Yan and Wei Liu and Yi Ding}, doi = {10.1016/j.engappai.2025.112491}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112491}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Constructing a robust short-text clustering model for contrastive learning based on optimized adaptive optimal transport for pseudo-label generation}, volume = {162}, year = {2025}, }
• (2025). Hybrid semantic-structural graph neural network for fault knowledge graph completion in railway operational equipment. EAAI, 162, 112490. (www)

  @article{ , abstract = {Incomplete fault knowledge graphs in railway operational equipment hinder effective fault diagnosis, prediction, and maintenance planning. This study addresses the challenge of completing fault knowledge graph by proposing a hybrid semantic-structural graph neural network (Hybrid S-GNN) that integrates both semantic and structural information for knowledge graph completion (KGC) in railway fault scenarios. The Hybrid S-GNN comprises four key modules: a semantic encoding module that enhances textual fault data representations through contextual enhancement and dynamic weight allocation; a structural encoding module that captures graph topology using multi-view structure encoding combining local aggregation, global path encoding, and relation-aware adjustment; a semantic-structural fusion module leveraging attention mechanisms to balance semantic and structural signals; and an optimization-prediction module employing margin-based ranking loss and context-aware negative sampling for accurate triple prediction. Experiments on real-world railway fault knowledge graphs demonstrate that Hybrid S-GNN achieves a Hits@10 of 80.5 % and mean reciprocal rank (MRR) of 0.640, outperforming state-of-the-art baselines by 5.8 % and 6.1 %, respectively. Ablation studies confirm the critical contributions of each module, validating the necessity of jointly modeling semantic and structural features. This work provides an effective solution to enhance railway fault knowledge graphs and paves the way for advanced fault management applications in railway operations.}, archive = {J_EAAI}, author = {Xiaorui Yang and Honghui Li and Yi Xu and Yunhao Deng and Yanhui Bai and Shufang Liu}, doi = {10.1016/j.engappai.2025.112490}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112490}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hybrid semantic-structural graph neural network for fault knowledge graph completion in railway operational equipment}, volume = {162}, year = {2025}, }
• (2025). Balancing efficiency and accuracy: Extreme gradient boosting and neural networks for near real-time brain deformation prediction in sports collisions. EAAI, 162, 112489. (www)

  @article{ , abstract = {Rapid head motion during sports collisions can cause traumatic brain injury. Head motion can be measured with instrumented mouthguards and fed into finite element (FE) models to predict brain strain, a measure of brain deformation and injury. Due to the computational cost of FE models, deep neural networks have been developed for near real-time prediction. However, they are not used in pitch-side assessments due to their complexity and reliance on full kinematic data, which cannot be reliably transmitted in real-time. We propose an extreme gradient boosting (XGBoost) model with simple input of two kinematic features. Its accuracy and efficiency were compared with two deep learning models: a multilayer perceptron (MLP) using 20 features, and a convolutional neural network (CNN) using entire kinematics. All models were trained on 1701 rugby impacts collected with mouthguards and simulated using the Imperial brain FE model. The XGBoost model predicted strain in key brain regions, while the deep learning models predicted whole-brain strain distributions. All models showed reasonable accuracy in predicting regional strain, with R 2 values 0.764–0.851 for XGBoost, 0.721–0.876 for MLP, and 0.744–0.887 for CNN. XGBoost required orders of magnitude fewer floating-point operations, and it used simple input that can be calculated on mouthguards and reliably transmitted in real-time. This study suggests that different models can be used at different stages of brain injury assessment. We hope that the XGBoost model proposed here will lower the barriers for adopting brain strain combined with instrumented mouthguards for pitch-side assessments from elite to grassroot collision sports.}, archive = {J_EAAI}, author = {Emily Yik Kwan Chan and Xiancheng Yu and Chen Qin and Mazdak Ghajari}, doi = {10.1016/j.engappai.2025.112489}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112489}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Balancing efficiency and accuracy: Extreme gradient boosting and neural networks for near real-time brain deformation prediction in sports collisions}, volume = {162}, year = {2025}, }
• (2025). Towards lensless image deblurring with prior-embedded implicit neural representations in the low-data regime. EAAI, 162, 112488. (www)

  @article{ , abstract = {The field of computational imaging has witnessed a promising paradigm shift with the emergence of untrained neural networks, offering novel solutions to inverse computational imaging problems. While existing techniques have demonstrated impressive results, they often operate either in the high-data regime, leveraging Generative Adversarial Networks (GANs) as image priors, or through untrained iterative reconstruction in a data-agnostic manner. This paper delves into lensless image reconstruction, a subset of computational imaging that replaces traditional lenses with computation, enabling the development of ultra-thin and lightweight imaging systems. To the best of our knowledge, we are the first to leverage implicit neural representations for lensless image deblurring, achieving reconstructions without the requirement of prior training. We perform prior-embedded untrained iterative optimization to enhance reconstruction performance and speed up convergence, effectively bridging the gap between the no-data and high-data regimes. Through a comprehensive comparative analysis of various untrained and low-shot methods, including under-parameterized non-convolutional techniques and domain-restricted low-shot approaches, we demonstrate that our method outperforms these alternatives by a significant margin. This work paves the way for the development of resource-constrained compact imaging systems, which can be applied in fields such as biomedical microscopy and embedded vision in low-data scenarios.}, archive = {J_EAAI}, author = {Abeer Banerjee and Sanjay Singh}, doi = {10.1016/j.engappai.2025.112488}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112488}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Towards lensless image deblurring with prior-embedded implicit neural representations in the low-data regime}, volume = {162}, year = {2025}, }
• (2025). Adaptive detection method for driver fatigue using facial multisource dynamic behavior fusion. EAAI, 162, 112482. (www)

  @article{ , abstract = {Driving while fatigued is a leading cause of traffic accidents. This study proposed an adaptive detection model to recognize driver fatigue based on the dynamic facial behavior information of drivers. First, drivers’ facial fatigue features were extracted to establish a general feature space, including pupil movement, eye state, and fatigue expression parameters. A differentiated feature space was then built based on individual drivers, taking into account the homogeneity, regularity, and individual variances in drivers' facial behavior at various states. A complete adaptive fatigue feature space was built by integrating the general feature space and differentiated feature space. Finally, a driver adaptive fatigue discrimination model was constructed to classify the general and adaptive fatigue feature space to detect driver fatigue states adaptively. A driver fatigue detection dataset from real scenarios had been established to validate the performance of the proposed model. Experimental results demonstrated that the proposed method significantly improved the detection accuracy of driver fatigue. In terms of artificial intelligence, this study contributes a novel adaptive feature space construction method based on multimodal dynamic feature fusion for facial fatigue recognition; in engineering application, it develops an adaptive driver fatigue detection system grounded in multimodal dynamic behaviors, which provides real-time alerts upon detecting driver fatigue and ensures driving safety.}, archive = {J_EAAI}, author = {Guoxin Zhang and Fei Yang and Xin Fang and Lili Wang and Lei Zhao and Chaoning Yu}, doi = {10.1016/j.engappai.2025.112482}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112482}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive detection method for driver fatigue using facial multisource dynamic behavior fusion}, volume = {162}, year = {2025}, }
• (2025). Thermoeconomic optimization of climate-adaptive solar and wind multi-generation systems using artificial intelligence and thermal energy recovery. EAAI, 162, 112481. (www)

  @article{ , abstract = {This study presents a hybrid multi-generation energy system designed to overcome solar intermittency while meeting the global demand for integrated delivery of electricity, water, cooling, and sustainable fuels in the transition to decarbonization. The engineering application integrates solar thermal and wind energy with a modified Brayton cycle, a Steam Rankine Cycle (SRC), and a Thermoelectric Generator (TEG) to simultaneously produce electricity, fresh water via Reverse Osmosis (RO), hydrogen and oxygen via Proton Exchange Membrane Electrolyzer (PEME), and cooling (via absorption chiller) within a unified optimization framework. The system was modeled using Engineering Equation Solver (EES) and optimized via Response Surface Methodology (RSM) based on 11 decision variables. To address the complexity of optimization, a second phase applied Artificial Intelligence (AI) techniques: Adaptive Boosting (AdaBoost) for predictive modelling and Particle Swarm Optimization (PSO) for global optimization. Under optimal conditions, the Response Surface Methodology yielded an exergy efficiency of 45.8 % with a cost rate of 576.76 United States Dollars per hour (USD/h), while AI reduced costs to 211.2 USD/h with a moderate efficiency trade-off. Simulation of the optimized configuration across eight diverse climates identified Quebec as most viable, generating 22,629.6 Megawatt-hours per year (MWh/year) of electricity and avoiding 4616.4 tons of Carbon Dioxide (CO 2 ) emissions annually. Integration of wind energy stabilizes solar variability, enhancing performance. AI contributes to optimizing complex interactions, nonlinear constraints, and multiple conflicting objectives. The methodology offers a scalable, generalizable framework for designing intelligent, climate-resilient infrastructures. Future research includes AI-enabled real-time control, experimental validation, and broader deployment strategies.}, archive = {J_EAAI}, author = {Ehsanolah Assareh and Nima Izadyar and Emad Tandis and Mehdi Khiadani and Amir shahavand and Neha Agarwal and Arian Gerami and Ahmed Rezk and Minkyu Kim and Reza Kord and Tahereh Pirhoushyaran and Mehdi Hosseinzadeh and Saleh Mobayen}, doi = {10.1016/j.engappai.2025.112481}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112481}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Thermoeconomic optimization of climate-adaptive solar and wind multi-generation systems using artificial intelligence and thermal energy recovery}, volume = {162}, year = {2025}, }
• (2025). A mamba-quantum attention transformer-convolutional network for automated pest and disease detection. EAAI, 162, 112480. (www)

  @article{ , abstract = {This study addresses critical challenges in apple leaf disease detection, where environmental interference (climate variations, lighting conditions, growth stages), symptom similarity across diseases, and phenotypic diversity within single diseases significantly impede accurate identification. To overcome these limitations, we propose the Mamba Quantum Attention Transformer-Graph Convolutional Network (MQAT-Transformer), a novel hybrid architecture that integrates quantum-enhanced attention mechanisms with dynamic graph learning. Firstly, the Mamba-Quantum X Attention mechanism (MQXA), inspired by quantum state modeling, optimizes visual feature extraction under complex environmental conditions. Secondly, the Dynamic Graph Convolution Feature Prioritization module (DGCF) adaptively resolves heterogeneous symptom manifestations by establishing multi-scale feature dependencies through learnable graph structures. This study conducted comprehensive experiments on the AppleLeaf9 Datasets (AppleLeaf9) proposed by Northwest A&F University, Apple disease leaf images Dataset (ADLI), and PlantVillage Dataset (PlantVillage) proposed by Pennsylvania State University. The results demonstrate that the proposed method achieves competitive performance across multiple evaluation metrics (e.g., accuracy, recall, and F1-score), confirming its generalization capability in cross-species scenarios. Finally, leveraging this network architecture, we developed a lightweight mobile application for farm leaf disease detection, offering a practical and user-friendly solution for crop health monitoring in agricultural settings.}, archive = {J_EAAI}, author = {Dong Tang and Zhihuan Liu and YiRui Zeng and Zhao Xu and Wendong Su}, doi = {10.1016/j.engappai.2025.112480}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112480}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A mamba-quantum attention transformer-convolutional network for automated pest and disease detection}, volume = {162}, year = {2025}, }
• (2025). Physics-driven prior learning-based deep unrolling for underwater image enhancement. EAAI, 162, 112472. (www)

  @article{ , abstract = {We propose a physics-driven prior learning-based algorithm unrolling approach for underwater image enhancement that leverages the advantages of both model- and learning-based approaches while overcoming their limitations. Model-based algorithms are theoretically robust because of prior knowledge of the underlying physics but may degrade image quality due to modeling inaccuracies. On the other hand, learning-based algorithms exhibit better adaptivity but inferior interpretability due to their black-box models and neglect of domain knowledge. In this work, we first formulate underwater image enhancement as a joint optimization problem with physics-based underwater-related priors and two learnable regularizers to compensate for modeling inaccuracies. Then, we solve the problem by reformulating it as a set of subproblems, which are then solved iteratively. Finally, we unroll the iterative algorithm into a deep neural network comprising a series of blocks, in which the optimization variables and regularizers are updated using closed-form solutions and learned deep neural networks, respectively. Experimental results on several datasets demonstrate that the proposed algorithm outperforms state-of-the-art underwater image enhancement algorithms on both quantitative and qualitative comparisons. The source code and pretrained models will be available at https://github.com/thithuypham/BLUE-Net .}, archive = {J_EAAI}, author = {Thuy Thi Pham and Hansung Yu and Truong Thanh Nhat Mai and Chul Lee}, doi = {10.1016/j.engappai.2025.112472}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112472}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Physics-driven prior learning-based deep unrolling for underwater image enhancement}, volume = {162}, year = {2025}, }
• (2025). Explore unicin milvus optimization-based scale invariant signal interference noise loss enabled cycle generative adversarial network for audio source separation. EAAI, 162, 112470. (www)

  @article{ , abstract = {Audio Source Separation remains the major research area as noisy audio signals lack information connectivity among users. Despite the numerous important efforts for audio source separation, there are still some challenges limiting the overall performance. Specifically, the existing methods face difficulty in working with multiple input sources and ignore the spectrum information of the signals, resulting in subpar performance. Hence, the research proposes the Explore Unicin Milvus Optimization-based Scale Invariant Signal interference Noise loss enabled Cycle Generative Adversarial Network (ExUnMO-S2NC-GAN) for addressing the challenges in the existing methods. The proposed approach utilizes the Cycle Generative Adversarial Network (C-GAN) architecture to carry out the transformation and restoration while preserving the significant details, resulting in achieving the most relevant audio source as the outcome. Specifically, the proposed model exploits the Scale Invariant Signal interference noise (S2N) loss function, improving the robustness against invariant to the signal scale and deformations of the signal. Besides, the Explore Unicin Milvus Optimization (ExUnMO) algorithm, harnessing the unique traits of Red Kite and Harris Hawk, is used for fine-tuning the hyperparameters of C-GAN, leading to improved performance. Moreover, the feature extraction with the spectral parameters added more advantages to the research model to work on more specific inputs. Extensive experiments demonstrates that theproposed model obtained high-efficiency outcomes in comparison with the state-of-the-art methods, which is evaluated with the error metrics attaining the Mean Absolute Error (MAE) of 1.79, and Mean Absolute Percentage Error (MAPE) of 3.22, whereas the signal quality metrics such as Peak Signal-to-Noise Ratio (PSNR), Signal-to-Interference Ratio (SIR) and Signal-to-Artifacts Ratio (SAR) achieved the high values of 54.81 dB (dB), 39.33 dB, and 40.56 dB respectively.}, archive = {J_EAAI}, author = {Baishakhi Dutta and Chandrakant J Gaikwad}, doi = {10.1016/j.engappai.2025.112470}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112470}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Explore unicin milvus optimization-based scale invariant signal interference noise loss enabled cycle generative adversarial network for audio source separation}, volume = {162}, year = {2025}, }
• (2025). A deep ensemble learning model for chinese spelling check. EAAI, 162, 112469. (www)

  @article{ , abstract = {Existing Chinese Spelling Check models are individual end-to-end models with different tendencies, which cannot fully cover all kinds of spelling errors and achieves advanced performance in every aspect. In this paper, a deep ensemble learning model for Chinese Spelling Check is designed, which captures the correct Chinese spelling answers proposed by most of the candidate ckeck models. Considering the phonological, visual and semantic characteristics of all candidate answers, it can even discover new correct answer which is not found by any of the candidate ckeck models. Specifically, in order to learn a hybrid representation of each input correction answer provided by each candidate ckeck model that includes phonological, visual, and semantic characteristics, a hybrid representation learner is designed and does not need to consider the compatibility of the candidate ckeck models. A deep ensemble correction network is designed to integrate all the hybrid representations and finds a final correction answer that considers all the useful information of all input correction answers. Moreover, based on the deep ensemble correction network, the deep ensemble learning model can be easily extended to involve new candidate ckeck models. Experimental results on the benchmark demonstrate that our proposed model largely outperforms any of its aggregated individual ckeck models.}, archive = {J_EAAI}, author = {Yaoyao Wu and Ruizhang Huang and Lina Ren and Ruina Bai}, doi = {10.1016/j.engappai.2025.112469}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112469}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A deep ensemble learning model for chinese spelling check}, volume = {162}, year = {2025}, }
• (2025). Efficient active flow control strategy for confined square cylinder wake using deep learning-based surrogate model and reinforcement learning. EAAI, 162, 112468. (www)

  @article{ , abstract = {This study introduces a deep learning surrogate model-based reinforcement learning (DL–MBRL) for active control of two-dimensional (2D) wake flow past a square cylinder confined between parallel walls using antiphase jets. In the training of this framework, a proximal policy optimisation (PPO) reinforcement learning agent alternates its interaction between a deep learning-based surrogate model (DL–SM) and a computational fluid dynamics (CFD) simulation to suppress wake vortex shedding, thereby significantly reducing computational costs. The DL–SM, built with a Transformer for temporal dynamics and a multiscale enhanced super-resolution generative adversarial network (MS–ESRGAN) for spatial reconstruction, is trained on 2D direct numerical simulation wake flow data to effectively and accurately emulate complex nonlinear flow behaviours. Compared to standard model-free reinforcement learning, the DL–MBRL approach reduces training time by about 50% while maintaining or improving wake stabilisation. Specifically, it achieves approximately a 98% reduction in shedding energy and a 95% reduction in the standard deviation of the lift coefficient, demonstrating strong suppression of vortex shedding. By leveraging the inherent stochasticity of DL–SM, DL–MBRL also addresses the nonzero mean lift coefficient issue observed in model-free methods, promoting more robust exploration. These results highlight the potential of the framework for extension to practical and industrial flow control problems.}, archive = {J_EAAI}, author = {Meng Zhang and Mustafa Z. Yousif and Minze Xu and Haifeng Zhou and Linqi Yu and Hee-Chang Lim}, doi = {10.1016/j.engappai.2025.112468}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112468}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Efficient active flow control strategy for confined square cylinder wake using deep learning-based surrogate model and reinforcement learning}, volume = {162}, year = {2025}, }
• (2025). Vessel behavior recognition method for preventing vessel-bridge collisions via adaptive video analytics. EAAI, 162, 112467. (www)

  @article{ , abstract = {This study presents a comprehensive end-to-end framework for recognizing vessel behaviors aimed at preventing vessel-bridge collisions in complex maritime environments. Current monitoring approaches that rely on Automatic Identification Systems or external tracking mechanisms often suffer from class imbalance, cross-domain variability, and limited capability to detect previously unseen high-risk behaviors. To address these challenges, the proposed framework directly analyzes video streams and introduces a standardized behavioral taxonomy, classifying vessel activities into eleven categories while incorporating temporal continuity and behavior transition modeling. A robust dataset construction pipeline is established, consisting of fixed-length frame sequences and mechanisms for cross-domain generalization. The framework integrates a spatio-temporal feature extraction module based on deformable convolution and multi-scale attention, coupled with a cross-instance mutual enhancement mechanism to capture domain-invariant representations. An open-set recognition strategy, grounded in class anchor clustering, enables accurate identification of previously unobserved high-risk behaviors. Furthermore, an adaptive frame sampling strategy dynamically adjusts sampling density around behavior transitions, enhancing recall and capturing infrequent events while minimizing computational cost. Extensive evaluations on both single-domain and multi-domain benchmark datasets, as well as real-world bridge video streams, demonstrate superior performance in terms of overall accuracy, F1-score, detection of rare behaviors, and recall compared with baseline methods. Ablation studies confirm the contribution of each component, and comparisons with open-set recognition methods underscore the practical utility of the proposed approach for anomaly detection. This framework provides a scalable, artificial intelligence-driven solution for vessel behavior recognition, anomaly detection, and cross-domain generalization, supporting intelligent monitoring and early warning in safety-critical maritime operations.}, archive = {J_EAAI}, author = {Woqin Luo and Daoxin Chen and Ye Xia and Dongming Feng}, doi = {10.1016/j.engappai.2025.112467}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112467}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Vessel behavior recognition method for preventing vessel-bridge collisions via adaptive video analytics}, volume = {162}, year = {2025}, }
• (2025). A secure architecture for intelligent internet of vehicle systems using advanced dimensionality reduction with deep ensemble models. EAAI, 162, 112466. (www)

  @article{ , abstract = {An intelligent transportation system (ITS) is an idea that provides a safe, protected, and smarter travelling experience to drivers. This visionary strategy aims to enable vehicles, pedestrian smartphones, roadside transportation frameworks, and additional devices to communicate with each other to offer secure and convenient services. Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communication in ITS enables the exchange of data like speed, position, and heading between vehicles and nearby networks. This improves data flow among infrastructure and vehicles, promoting road safety, traffic efficiency, reduced emissions, and ITS development. Artificial intelligence (AI) now presents smart solutions for addressing these threats effectively. This paper presents a Secure Architecture for Intelligent Vehicle Systems Using Integrated Ensemble Model and Advanced Dimensionality Reduction (SAIVS-IEMADR) model. Initially, the data pre-processing is performed to transform and organize raw data into a suitable format. The feature selection (FS) process uses five methods: Consistency, Correlation Feature Selection (CFS), information theory (InfoGain), Relief for Features (Relief-F), and Fast-Correlation Based Filter (FCBF) to identify the most relevant and informative features. Finally, an ensemble model comprising the temporal convolutional network (TCN), deep belief network (DBN), and stacked sparse auto-encoder (SSAE) is employed for classification. The SAIVS-IEMADR approach was evaluated on the Network Security Laboratory – Knowledge Discovery in Databases (NSLKDD) and Canadian Institute for Canadian Institute for Cybersecurity Intrusion Detection System (CICIDS)-2017 datasets, achieving superior accuracy of 99.55 % and 99.32 % compared to existing methods.}, archive = {J_EAAI}, author = {Amal K. Alkhalifa and Mashael M. Asiri and Reham Al-Dayil and Ahmed Alsayat and Mashail N. Alkhomsan and Mohammed A. AlAqil and Shouki A. Ebad and Mohammed Mujib Alshahrani}, doi = {10.1016/j.engappai.2025.112466}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112466}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A secure architecture for intelligent internet of vehicle systems using advanced dimensionality reduction with deep ensemble models}, volume = {162}, year = {2025}, }
• (2025). A constitutive framework for non-isothermal plasticity through micro-mechanism informed artificial neural networks. EAAI, 162, 112465. (www)

  @article{ , abstract = {Non-isothermal plastic deformation creates modeling challenges for constitutive model development. Current constitutive models challenge to capture the coupled microstructural changes that occur when temperature varies during processing. The Micro-Mechanism Informed Artificial Neural Network (MMIANN) framework was developed to address these limitations using artificial intelligence (AI) and machine learning (ML) techniques. The MMIANN framework combines physics-based evolution equations for key metallurgical processes—dislocation density, grain boundary migration, and precipitation kinetics—with neural network predictions. These micro-mechanisms operate as internal state variables that guide the network's material behavior predictions. The architecture uses parallel physics-based and neural pathways, blended through adaptive coefficients. Thermodynamic constraints maintain consistency through penalty-based enforcement of the Clausius-Duhem inequality. The model was trained and validated using experimental data from simultaneous cooling and tensile deformation of AA7075 aluminum alloy. The tests replicated industrial hot forming conditions with cooling rates from 25 to 75 °C per second (°C/s). This experimental approach captures the thermal-mechanical coupling that drives microstructural evolution in practice. MMIANN achieved correlation coefficients exceeding 0.96 for stress, temperature, and grain size predictions. The framework captures thermomechanical regimes. Processing maps generated by the AI model link process parameters to microstructural outcomes. The analysis reveals an optimal processing window (40–55 °C/s cooling, 0.15–0.25 strain ( ε )) and identifies three regimes where different strengthening mechanisms dominate. By integrating metallurgical science with machine learning, this framework provides a practical tool for non-isothermal manufacturing processes. The approach bridges microstructural understanding with process control for thermal-mechanical operations through the application of artificial intelligence.}, archive = {J_EAAI}, author = {Yo-Lun Yang and Tsai-Fu Chung and Chia-Hung Liao and Liang-Yu Chen and Hsing-Yu Wu and Uthayakumar Marimuthu and Arumugaprabu Veerasimman and Sundarakannan Rajendran and Vigneshwaran Shanmugam}, doi = {10.1016/j.engappai.2025.112465}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112465}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A constitutive framework for non-isothermal plasticity through micro-mechanism informed artificial neural networks}, volume = {162}, year = {2025}, }
• (2025). Learnable patchmatch and self-teaching for multi-frame depth estimation in monocular endoscopy. EAAI, 162, 112463. (www)

  @article{ , abstract = {This work delves into unsupervised monocular depth estimation in endoscopy, which leverages adjacent frames to establish supervisory signals during the training phase. For many clinical applications, e.g. , surgical navigation, temporally correlated frames are also available at test time. However, most existing monocular methods struggle to make effective use of temporal information during both training and inference, primarily due to the inherent challenges of endoscopic imagery, including low- or homogeneous-texture regions and brightness fluctuations between frames. To fully exploit the temporal information in endoscopic scenes, we propose a novel unsupervised multi-frame monocular depth estimation model. The proposed model integrates a learnable patchmatch module to adaptively increase the discriminative ability in regions with low or homogeneous textures, and enforces cross-teaching and self-teaching consistencies to provide efficacious regularizations towards brightness fluctuations. Furthermore, as a byproduct of the self-teaching paradigm, the proposed model is able to improve the depth predictions when more frames are input at test time. We conduct detailed experiments on multiple datasets, and the experimental results indicate that the proposed method exceeds prior state-of-the-art competitors. The source code and trained models will be publicly available at https://github.com/ShuweiShao/FrameDepth .}, archive = {J_EAAI}, author = {Shuwei Shao and Zhongcai Pei and Weihai Chen and Xingming Wu and Zhong Liu}, doi = {10.1016/j.engappai.2025.112463}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112463}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Learnable patchmatch and self-teaching for multi-frame depth estimation in monocular endoscopy}, volume = {162}, year = {2025}, }
• (2025). Hypergraph induced semi-supervised orthogonal nonnegative matrix factorization with label and constraint propagation. EAAI, 162, 112462. (www)

  @article{ , abstract = {As a popular technology of artificial intelligence, nonnegative matrix factorization (NMF) aims at clustering and finding the differentially expressed features of each cluster. However, for complex high-dimensional sample data, it is still a challenge to design more appropriate NMF optimization models and develop more efficient algorithms to solve this model in view of enhanced theoretical properties and numerical performance. In this paper, a novel NMF optimization model with regularization is proposed such that the NMF is performed by a semi-supervised approach, as well as incorporating the strategies of hypergraph induced label propagation and constraint propagation. Specifically, different from existing NMF methods, the hypergraph structure underlying the data, together with the simple graph information, is employed to guide the pairwise constraint propagation in our built model. In recognition of sample similarity, a dataset-adaptive strategy is proposed to update the weight matrix of the graphs. By adding dual orthogonality on the factor matrices in the objective function, interpretability and feature independence of the built model are enhanced. Then, an algorithm is developed to efficiently solve this complicated model. Theoretically, it is proved that the developed algorithms are well defined and convergent. Numerically, extensive tests on the proposed model and algorithm are performed, which validate that they outperform the state-of-the-art ones in terms of different metrics of evaluating clustering performance when they are applied into solution of the problems from eight public datasets.}, archive = {J_EAAI}, author = {Jie Guo and Ting Li and Jialu Liu and Zhong Wan}, doi = {10.1016/j.engappai.2025.112462}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112462}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hypergraph induced semi-supervised orthogonal nonnegative matrix factorization with label and constraint propagation}, volume = {162}, year = {2025}, }
• (2025). A novel condition monitoring approach using hybrid lightweighted adaptive models for complex machinery. EAAI, 162, 112461. (www)

  @article{ , abstract = {Condition monitoring of complex industrial systems is critical for ensuring operational reliability. Data-driven methods using artificial intelligence have advanced anomaly detection (AD) and fault diagnosis (FD), but existing approaches often treat them separately, focus on known faults, and struggle with previously unseen or rare conditions in multi-modal scenarios. This study proposes a novel condition monitoring framework that integrates AD and FD within a distributed architecture. Lightweight models—including kernel principal component analysis, support vector machines, and one-dimensional convolutional neural networks—enable efficient and scalable processing. A multilevel information fusion strategy ensures consistent detection and diagnosis while facilitating the isolation of previously unknown faults. Module test results demonstrate the effectiveness and robustness of the proposed feature extraction and adaptive modeling approaches. The overall test results for previously unknown faults vary across channels and modules. For samples with misalignment and inner blade wear, channel-level detection accuracy ranges from 0.007 to 0.989, with unknown recognition rates up to 0.933 and diagnosis probabilities from 0.508 to 0.933. For strong misalignment and fan-end inner race faults, nearly all channels achieve 100 % detection accuracy, with some diagnosis probabilities above 0.9, while unknown recognition remains minimal (mostly below 0.05). Importantly, the proposed framework integrates detection and diagnostic outputs across channels, effectively mapping previously unseen faults to similar known categories or to an unknown category. Overall, the proposed framework offers a referenced solution for condition monitoring of industrial systems like pumps, turbines, and compressors, and lays the foundation for future improvements incorporating domain knowledge and model-driven interpretability.}, archive = {J_EAAI}, author = {Yingqian Liu and Rongyong Zhang and Luigi Grossi and Zhipin Ye and Huairui Li and Rongsheng Zhu and Qiang Fu}, doi = {10.1016/j.engappai.2025.112461}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112461}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel condition monitoring approach using hybrid lightweighted adaptive models for complex machinery}, volume = {162}, year = {2025}, }
• (2025). A fast and stable framework for generative adversarial imitation learning. EAAI, 162, 112460. (www)

  @article{ , abstract = {Applying reinforcement learning to complex and high-dimensional tasks encounters challenges, including the formulation of suitable reward functions, achieving sample efficiency, and developing effective exploration strategies. Adversarial imitation learning techniques address these issues by employing generative adversarial networks (GANs) to capture temporal dependencies and mitigate compounding errors. These methods exhibit enhanced efficiency relative to behavioral cloning and require fewer expert samples. Nonetheless, their complex min-max problem results in sample inefficiency and struggles with challenges such as mode collapse and training instability. This paper provides a comprehensive approach that adeptly addresses these challenges. The proposed method's design comprises three main steps. First, it uses the off-policy Twin Delayed Deep Deterministic (TD3) algorithm to enhance sample efficiency and accelerate learning. In the second step, a novel reward function based on energy-based GANs and deep regret analytic GANs is developed, which alleviates mode collapse and enhances training stability. Finally, we suggest several improvements, including the use of a pre-trained discriminator and mixed batches, to achieve a faster and more stable algorithm. The evaluation findings on continuous control tasks demonstrate that our method not only matches state-of-the-art performance but also surpasses it in both sample efficiency and stability. It achieves convergence with far fewer iterations than the compared methods.}, archive = {J_EAAI}, author = {Fateme Shahabi-Nejad and Mohammad Mehdi Ebadzadeh}, doi = {10.1016/j.engappai.2025.112460}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112460}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A fast and stable framework for generative adversarial imitation learning}, volume = {162}, year = {2025}, }
• (2025). A transfer learning method of collaborating random walk and adaptive instance normalization for inscription image denoising. EAAI, 162, 112458. (www)

  @article{ , abstract = {Mess noise hinders reading and understanding of inscriptions in images. For image restoration from noise-corrupted images, existing network-learning-based methods can construct an excellent model to generate noise patterns. However, the performance of such models is degraded owing to the lack of high-quality training data and the complex noise pattern in inscription images, e.g., mixed noise with multiple levels. Herein, we first propose a novel noise generation model that can produce more realistic synthetic noise images using the random walk algorithm. Then, we propose an explainable inscription image denoising network using a variational inference model, where the joint distribution of clean-noise image pairs is approximated in a dual adversarial manner. The proposed network exhibits improved generalizability and adaptability to different noise characteristics using an estimated noise map and adaptive instance normalization. Finally, we introduce a transfer learning scheme to migrate the network learned from the synthetic noise image domain to a real-inscription image domain with a limited number of real-inscription images. The proposed method outperforms state-of-the-art methods.}, archive = {J_EAAI}, author = {Erhu Zhang and Yunjing Liu and Guangfeng Lin and Jinghong Duan}, doi = {10.1016/j.engappai.2025.112458}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112458}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A transfer learning method of collaborating random walk and adaptive instance normalization for inscription image denoising}, volume = {162}, year = {2025}, }
• (2025). Deep learning-based time series forecasting for bearing remaining useful life: Recent advances, hybrid architectures, and targeted enhancements. EAAI, 162, 112457. (www)

  @article{ , abstract = {This paper presents a systematic review of deep learning(DL)-based time series forecasting hybrid model for bearing RUL prediction by evaluating these studies from 2020 to 2025, addressing the gap in analyzing cross-model collaboration strategies. Firstly, the general workflow and modeling framework for remaining useful life (RUL) prediction were introduced. Secondly, a comparative analysis of time dependence capture mechanisms -including Recurrent Neural Networks(RNN) and their variants and Temporal Convolutional Network(TCN), and Transformer-was conducted. The findings reveal a shift from local sequential modeling to globally contextualized frameworks, enhanced by self-supervised learning and multimodal fusion. These developments underscore the importance of aligning model selection with specific task objectives and balancing multiple dimensions to achieve optimal performance in complex industrial scenarios. Thirdly, to address common challenges in RUL prediction, such as noisy data, sparse features, high model complexity, and limited interpretability, this paper explores performance optimization strategies from the perspectives of data quality, architectural design, and decision transparency. Finally, a comprehensive review was made of the current challenges: unclear applicability of attention, limited interpretability, deployment difficulties under resource constraints, difficulties in real-time prediction, and weak generalization in a data-scarce environment. To address these issues, future research should explore more trustworthy attention evaluation, interpretable architectures, lightweight design, continuous iteration for real-time prediction, and leveraging embedding other models to complement the lack of generalization. Overall, this work provides a structured perspective to support the development of the next generation of intelligent prediction systems.}, archive = {J_EAAI}, author = {Zhenyu Tang and Dalian Yang and Linrong Tan and Liying Zeng}, doi = {10.1016/j.engappai.2025.112457}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112457}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep learning-based time series forecasting for bearing remaining useful life: Recent advances, hybrid architectures, and targeted enhancements}, volume = {162}, year = {2025}, }
• (2025). Finite element-integrated neural network framework for spatial modal prediction in machine tool structures. EAAI, 162, 112456. (www)

  @article{ , abstract = {Accurate prediction of position-dependent structural dynamics in (CNC) machine tools is critical for ensuring machining precision, avoiding resonance, and enabling real-time health monitoring. This study proposes an integrated framework combining finite element analysis (FEA), experimental modal testing, and artificial neural networks (ANNs) to model and predict position-dependent dynamics across the machine workspace. A validated FEA model of a high-rigidity vertical machining centre is developed and correlated with experimental modal analysis using a PCB 086D20 impact hammer and tri-axial accelerometer. Natural frequencies and mode shapes are extracted and compared, showing deviation under 10 %, confirming model fidelity. To capture position-dependent dynamics, modal analysis was performed at 27 spatial locations, revealing significant frequency variation across planes, indicating localized compliance zones. A multilayer ANN is trained on the modal dataset to predict frequencies based on spatial coordinates, achieving R 2 values above 0.99. The proposed hybrid approach enables real-time estimation of structural dynamics, reducing the need for repeated testing and supporting intelligent control strategies in large-format CNC systems. This work contributes a predictive foundation for dynamic stability optimization, resonance avoidance, and digital twin development in precision machining applications.}, archive = {J_EAAI}, author = {Aman Ullah and Tzu-Chi Chan and Jun-Fa Huang and Shinn-Liang Chang}, doi = {10.1016/j.engappai.2025.112456}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112456}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Finite element-integrated neural network framework for spatial modal prediction in machine tool structures}, volume = {162}, year = {2025}, }
• (2025). End-to-end go (Weiqi) game record reconstruction from live broadcast videos. EAAI, 162, 112455. (www)

  @article{ , abstract = {Manual transcription of Go (Weiqi) game records from broadcast videos creates a major bottleneck for building training data in artificial intelligence (AI). We present an end-to-end system that reconstructs complete smart game format (SGF) records from professional tournament broadcasts. The pipeline integrates four modules: (i) video-to-frame sampling; (ii) a board detector based on a detection transformer (DETR), optimized with a size-aware loss to reliably localize both main and commentary boards; (iii) a stone classifier that uses knowledge distillation from a vision transformer (ViT) teacher to an efficient grid-based student for full-board inference; and (iv) a temporal reconstruction algorithm for occlusion recovery and move-sequence consistency. When evaluated under a zero-error-tolerance video-level protocol – where a video is counted correctly only if all sampled frames match the reference – the system achieves 82.56% accuracy on 86 real tournament videos (356 h). Component analyses reveal high board localization quality with mean average precision (mAP50-95) reaching 0.99; near-teacher board-state recognition with a 67 × speedup (377.5 frames per second, FPS) and F1 score 0.988; and a 70.23% reduction of occlusion-related misdetections. Compared with you only look once (YOLO) and faster region-based convolutional neural network (Faster R-CNN) baselines, our design improves small-board recall and end-to-end robustness in dual-board and occluded settings. The system outputs SGF records suitable for large-scale dataset construction, AI-assisted analysis, education, and digital preservation, and the approach can be generalized to other grid-structured board games with minor adaptations.}, archive = {J_EAAI}, author = {Chih-Lin Lin and Hsia-Hung Ou and Lung Hung Chen and Chih-En Kuo}, doi = {10.1016/j.engappai.2025.112455}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112455}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {End-to-end go (Weiqi) game record reconstruction from live broadcast videos}, volume = {162}, year = {2025}, }
• (2025). Detecting code paraphrased by large language models using coding style features. EAAI, 162, 112454. (www)

  @article{ , abstract = {Recent progress in large language models (LLMs) for code generation has raised serious concerns about intellectual property protection. Malicious users can exploit LLMs to produce paraphrased versions of proprietary code that closely resemble the original. While the potential for LLM-assisted code paraphrasing continues to grow, research on detecting it remains limited, underscoring an urgent need for a detection system. We respond to this need by proposing two tasks. The first task is to detect whether code generated by an LLM is a paraphrased version of original human-written code. The second task is to identify which LLM is used to paraphrase the original code. For these tasks, we construct a dataset consisting of pairs of human-written code and LLM-paraphrased code using various LLMs. We statistically confirm significant differences in the coding styles of human-written and LLM-paraphrased code, particularly in terms of naming consistency, code structure, and readability. Based on these findings, we develop a detection method that identifies paraphrase relationships between human-written and LLM-generated code, and discover which LLM is used for the paraphrasing. Our detection method outperforms the best baselines in two tasks, improving F1 scores by 2.64% and 15.17% while achieving speedups of 1,343x and 213x, respectively.}, archive = {J_EAAI}, author = {Shinwoo Park and Hyundong Jin and Jeong-won Cha and Yo-Sub Han}, doi = {10.1016/j.engappai.2025.112454}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112454}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Detecting code paraphrased by large language models using coding style features}, volume = {162}, year = {2025}, }
• (2025). Cooperative neural networks for inverse design: Integrating denoising autoencoder and surrogate model for partial design variable imputation. EAAI, 162, 112453. (www)

  @article{ , abstract = {Data-driven inverse design is an engineering approach where target performance criteria are specified upfront, leading to the derivation of design solutions that meet these criteria. While recent research focuses on generating complete design solutions using generative models, these approaches struggle with partial design variables and constraints that predetermine certain variables. Additionally, generative models are data-intensive and prone to overfitting with limited datasets. To address these limitations, this paper proposes a Cooperative Neural Network architecture comprising two key components: the Imputation Model and the Surrogate Model. These components collaborate to optimize design solutions while adhering to predefined performance criteria. The framework’s effectiveness is demonstrated through a case study on Glass Run Channel (GRC) designs from a Korean automotive manufacturer. Results show the architecture proficiently imputes undetermined variables and ensures the designs meet desired performance metrics, achieving Mean Squared Error (MSE) reductions of up to 98 % and R-squared values of 0.997–0.999 in initial tests. It remains robust in diverse scenarios, achieving up to 95.65 % MSE reduction and R-squared values of 0.995–0.999 for cases with the most undetermined variables, and up to 94.68 % MSE reduction with R-squared values of 0.983–0.995 for the smallest training datasets. This framework reduces design cycle times and enhances engineering design efficiency, offering a robust solution to limitations in traditional methods reliant on physical prototyping and iterative testing.}, archive = {J_EAAI}, author = {Agung Nugraha and Hyerin Kwon and Gyeongho Park and Jihwan Lee}, doi = {10.1016/j.engappai.2025.112453}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112453}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Cooperative neural networks for inverse design: Integrating denoising autoencoder and surrogate model for partial design variable imputation}, volume = {162}, year = {2025}, }
• (2025). Knowledge distillation via teacher-modeled sample relationships for skin cancer diagnosis. EAAI, 162, 112452. (www)

  @article{ , abstract = {In skin cancer diagnosis, knowledge distillation can achieve model compression, which enables complex deep learning tasks to run efficiently on devices with limited resources. However, existing studies have largely overlooked explicit modeling of sample relationships during teacher model training, particularly in the context of class imbalance. Directly transferring such unrefined relational knowledge not only fails to address the bias toward majority classes but also leads to suboptimal performance for minority classes. To address this gap, we propose a novel approach that integrates a sample relationship module into the teacher model to learn relational knowledge under class imbalance, which is then transferred to the student model during distillation. Specifically, the teacher transfers two key types of knowledge: traditional logits and inter-sample relationships. During training, the student is optimized through two losses: a standard logit-matching loss to mimic classification knowledge, and relational consistency losses to enforce alignment between the student’s and teacher’s predicted inter-sample relationships to address class imbalance. We conducted extensive experiments on large-scale public skin lesion classification datasets, including ISIC2019 and HAM10000, and the results demonstrate that our method significantly outperforms current state-of-the-art approaches.}, archive = {J_EAAI}, author = {Peng Liu and Wenhua Qian and Shan Tang}, doi = {10.1016/j.engappai.2025.112452}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112452}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Knowledge distillation via teacher-modeled sample relationships for skin cancer diagnosis}, volume = {162}, year = {2025}, }
• (2025). Region-based weighting-and-enhancement network with adaptive class weighting loss for postoperative inguinal hernia prediction. EAAI, 162, 112451. (www)

  @article{ , abstract = {Postoperative inguinal hernia (PIH) is a common complication after radical prostatectomy, subsequently leading to multiple potential risks (e.g., cardiovascular and cerebrovascular accidents) and increased surgical costs due to re-surgical reparation. Magnetic resonance imaging (MRI) examination is a widely used procedure before radical prostatectomy, which can investigate the muscle structures of the abdominal wall (MSAW). Recently, clinical studies have indicated that clinical parameters (e.g., thickness and width of the external oblique muscle) of MSAW are strongly related to PIH. However, automated MRI-based PIH prediction based on deep neural networks has not been studied previously. Motivated by these observations, we propose a novel region-based weighting-and-enhancement network to predict PIH before radical prostatectomy based on MRI images automatically. Specifically, we employ the well-designed Region Weighting-and-Enhancement module to capture informative context representations through region weighting and regional context enhancement, by fully leveraging the potential of clinical MSAW priori. Additionally, this paper designs an effective adaptive class weighting loss to emphasize or suppress the samples with varying levels of significance to further boost the PIH prediction performance. The extensive experiments on a clinical MRI-PIH dataset and one publicly available MRI dataset manifest the superiority of our proposed methods over state-of-the-art deep neural networks and advanced loss methods.}, archive = {J_EAAI}, author = {Jiawei Zhang and Lisheng Wu and Qiang Fang and Weidong Yu and Zhengyu Hu and Fengyun Zhang and Cheng Yang and Xiaoqing Zhang}, doi = {10.1016/j.engappai.2025.112451}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112451}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Region-based weighting-and-enhancement network with adaptive class weighting loss for postoperative inguinal hernia prediction}, volume = {162}, year = {2025}, }
• (2025). Predicting thermal transport of blood-based penta-hybrid nanofluid in fin geometries using deep neural networks and finite difference approach. EAAI, 162, 112450. (www)

  @article{ , abstract = {Nanofluids have garnered significant research interest due to their enhanced heat transfer and thermal characteristics. A novel hybrid nanofluid has exhibited exceptional thermal properties, combining five nanoparticles of uniform shapes with a base fluid, such as blood. This study investigates the influence of fin thickness, varying with length, considering the implications of internal heat production, convection, and thermal radiation processes in rectangular, convex, and triangular fin descriptions. Wet scenarios are interpreted to evaluate differences in thermal energy dynamics for fin shapes like Rectangular, Convex and Triangular. Darcy's model is employed to account for the material's porous nature. A finite difference scheme, implemented using Partial Differential Equation solver (PDSolve) in Maple (2024), provides graphical insights into fin effectiveness and thermal steady-state responses across various parameters. Incorporating Penta hybrid nanofluids enhances fin performance, with rectangular fins' Nusselt numbers (up to 1.936) proving more efficient, delivering faster thermal responses than triangular fins and convex fins. Further, using the Adam Optimisation algorithm, Convolutional Neural Networks were used to validate the current model. It was observed that these networks could accurately forecast the truth values, and the two findings matched, as indicated in Table 3 As a potential biological application, this research offers insight into optimising cooling systems for biomedical devices, such as heat exchangers in artificial organs.}, archive = {J_EAAI}, author = {Maddina Dinesh Kumar and Nehad Ali Shah and Dharmaiah Gurram and Se-Jin Yook}, doi = {10.1016/j.engappai.2025.112450}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112450}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Predicting thermal transport of blood-based penta-hybrid nanofluid in fin geometries using deep neural networks and finite difference approach}, volume = {162}, year = {2025}, }
• (2025). Research on the application of attention mechanism based multi-model fusion in food recommendation platforms. EAAI, 162, 112449. (www)

  @article{ , abstract = {Smartphone-based food ordering has greatly enhanced convenience in daily life, and the rise of recommendation systems has transformed the functionality and user experience of food delivery applications. Innovations in recommendation algorithms and models have significantly improved the efficiency of food, merchant, and advertisement recommendations on food platforms, leading to higher transaction rates and greater user satisfaction. To further enhance recommendation efficiency, this study introduces a novel multi-model fusion recommendation architecture based on the multi-head self-attention mechanism, utilizing a two-tier structure. The first-tier model (the attention-based homogeneous AutoInt model) acts as a teacher to guide the training of the second-tier Transformer model. This hierarchical approach integrates multiple models through knowledge distillation, significantly improving the accuracy of the recommendation system. The complexity and performance of the proposed architecture were analyzed and applied in a production environment. Testing on a private dataset reveals that the proposed multi-model fusion recommendation architecture significantly enhances recommendation performance across various food platform scenarios, achieving an accuracy of 0.7643, recall of 0.8262, and an F1 score of 0.7936. These results surpass the performance of current state-of-the-art models. Therefore, the proposed architecture is not only highly applicable to food recommendation systems but also has broad applicability in other fields such as retail and entertainment.}, archive = {J_EAAI}, author = {Linchao Zhang and Lei Hang}, doi = {10.1016/j.engappai.2025.112449}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112449}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Research on the application of attention mechanism based multi-model fusion in food recommendation platforms}, volume = {162}, year = {2025}, }
• (2025). Identifying fake reviews for refund purposes: Evaluating the effectiveness of a transfer-learning model against emerging large language models. EAAI, 162, 112448. (www)

  @article{ , abstract = {Background: Recently, dishonest sellers are using social platforms to advertise products that can be purchased for free through a refund mechanism, which is based on the writing of five-star fake reviews. The aim is to increase product visibility by influencing their ranking compared to similar products. This mechanism is leading to a significant distortion of e-commerce platforms, eroding trust among customers and sellers. Objective: In this paper, we address the problem of identifying fake reviews, aiming to provide an approach for mitigating fraudulent practices that compromise the integrity and transparency of e-commerce platforms. Methods: We propose a supervised model tailored for identifying fake reviews for refund purposes and compare its performance with some of the most recent generative models. Since, to the best of our knowledge, no datasets exist in the literature suitable for fake review identification in the process of Purchasing, Requesting reviews, and Refunding a product, we first proposed a new dataset of fake and genuine reviews from Amazon, collected with the help of a domain expert. Then, we defined five other new datasets containing reviews automatically generated by language models. To interact with these models, we designed new prompt approaches specifically tailored to our goal, which exploit the iterative refinement behind these models for improving classification results. Results: Experimental results demonstrated the effectiveness of the supervised model in detecting both types of fake reviews, outperforming state-of-the-art models with improvements ranging from 0.23 to 0.70 in terms of accuracy, precision, and recall.}, archive = {J_EAAI}, author = {Loredana Caruccio and Gaetano Cimino and Stefano Cirillo and Vincenzo Deufemia and Giuseppe Polese and Giandomenico Solimando}, doi = {10.1016/j.engappai.2025.112448}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112448}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Identifying fake reviews for refund purposes: Evaluating the effectiveness of a transfer-learning model against emerging large language models}, volume = {162}, year = {2025}, }
• (2025). Enhanced multi-modal emotion recognition using the feature level fusion. EAAI, 162, 112447. (www)

  @article{ , abstract = {Multi-modal human emotion recognition is a complex process of synthesizing information from various modalities to calculate emotion states. This field faces several challenges: (1) Acoustic is an essential component of emotion expression, but it often underperforms compared to visual and text in emotion recognition. (2) Capturing the feature interaction among different modalities is usually complex. (3) Processing high-definition videos can significantly reduce the efficiency of visual analysis. In this study, we presented a learning architecture designed to recognize human emotions effectively. For the first challenge, we implemented a multi-level acoustic encoder (MLAE) that enhances the extraction of acoustic information to improve the acoustic contribution in multi-modal emotion recognition. Facing the second challenge, we introduced the cross-attention block module, which adeptly captures the inter-modal interactions. To address the third challenge, we adopted the re-parameterized visual geometry group network (RepVGG) as the visual feature encoder, employing its multi-branch learning and single-branch reasoning structure to maintain high reasoning efficiency. Our model has demonstrated the state-of-the-art performance of the interactive emotional dyadic motion capture (IEMOCAP) dataset and the multi-modal opinion sentiment and emotion intensity of the Carnegie Mellon University (CMU-MOSEI) dataset.}, archive = {J_EAAI}, author = {Aziguli Wulamu and Yuheng Wu and Xin Liu and Yao Zhang and Jinghan Xu and Yang Zhang}, doi = {10.1016/j.engappai.2025.112447}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112447}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhanced multi-modal emotion recognition using the feature level fusion}, volume = {162}, year = {2025}, }
• (2025). Deep adaptive wavelet autoencoder with mutually independent empirical cumulative distribution for unsupervised motor anomaly detection. EAAI, 162, 112446. (www)

  @article{ , abstract = {—Permanent magnet synchronous motors (PMSMs) are widely used in industrial applications but remain vulnerable to stator faults, such as inter-coil and inter-turn short circuits. Although recent deep learning-based fault detection methods have shown promise, they typically rely on large volumes of labelled fault data for training. To address this limitation, this paper proposes a novel unsupervised fault detection framework, termed Deep Adaptive Wavelet Autoencoder (DAWA) with Mutually Independent Empirical Cumulative Distribution (MIECD), specifically designed for PMSM fault detection. DAWA utilizes convolutional neural networks to learn adaptive wavelet filters through fast discrete wavelet transform, allowing for fully learnable, threshold-free extraction of fine-grained signal patterns. The resulting latent features are then mapped by MIECD into a mutually independent space via independent component analysis (ICA). Without assuming any prior data distribution, MIECD estimates empirical cumulative distributions (ECDs), computes tail probabilities across dimensions, and aggregates them into a unified anomaly score. Experimental results on motor vibration datasets demonstrate the effectiveness of the proposed method, showing average accuracy improvements of 15.85 % for Interturn and 15.16 % for Intercoil fault detection compared to conventional data-driven baselines across various operating conditions.}, archive = {J_EAAI}, author = {Pinze Ren and Ning Zhu and Dandan Peng and Liyuan Ren and Huan Wang}, doi = {10.1016/j.engappai.2025.112446}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112446}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep adaptive wavelet autoencoder with mutually independent empirical cumulative distribution for unsupervised motor anomaly detection}, volume = {162}, year = {2025}, }
• (2025). Dynamic vision-based machine vibration sensing and fault diagnosis with signal alignment and feature clustering. EAAI, 162, 112445. (www)

  @article{ , abstract = {In recent years, intelligent fault diagnosis methods for machines have been widely developed. The contact accelerometer has been popularly used with high measurement accuracy. However, in many industrial applications, non-contact sensors are often preferred due to practical constraints. The event camera is a novel bio-visual sensing technology for asynchronously capturing pixel-wise changes in brightness. It has various advantages including high measurement rate, exceptional resolution, wide dynamic range, etc., which thus has promising prospects in non-contact monitoring and fault diagnostic tasks. Despite these advantages, the high complexity of the dynamic vision data from the event cameras poses significant processing challenges, and the extraction of machine vibration information is of great difficulties. To address these challenges, this paper proposes a novel dynamic vision-based machine vibration sensing and fault diagnosis method. First, the dynamic vision data is reconstructed into event frame sequences. Next, a deep neural network is proposed to extract the micro-vibration information with feature clustering for enhancing model robustness. A signal alignment method is further proposed where the contact sensing data are used as a reference for optimizing model performance. Finally, the intelligent fault diagnosis is implemented with the estimated vibration data. Experimental validations are conducted with real rotating machine data, which demonstrate the promising applicability of the proposed method in non-contact machine vibration sensing and fault diagnosis.}, archive = {J_EAAI}, author = {Ruiyi Guang and Xiang Li and Yaguo Lei and Bin Yang and Naipeng Li}, doi = {10.1016/j.engappai.2025.112445}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112445}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dynamic vision-based machine vibration sensing and fault diagnosis with signal alignment and feature clustering}, volume = {162}, year = {2025}, }
• (2025). A reliable degradation prediction method for proton exchange membrane fuel cells based on uncertainty bayesian self-attention. EAAI, 162, 112444. (www)

  @article{ , abstract = {Health state prediction of Proton Exchange Membrane Fuel Cells (PEMFCs) is a critical technology to ensure their long-term reliable operation. Prediction accuracy directly influences the effectiveness of maintenance strategies and risk management. However, existing PEMFC degradation prediction methods based on Recurrent Neural Networks (RNNs) or Transformer architectures mostly focus on point estimation while neglecting uncertainty quantification. This limitation makes it difficult to assess the confidence level of predictions in practical engineering applications, reducing the models' reliability in decision support. To address this issue, this paper proposes a novel Bayesian Patch Time Series Transformer (B-PatchTST) method. By deeply integrating Bayesian variational inference with time series patch modeling, the method enables probabilistic prediction of PEMFC degradation trajectories and disentangled analysis of uncertainty sources. Unlike traditional Bayesian Neural Networks (BNNs) that primarily apply Bayesian modeling to fully connected layers, B-PatchTST introduces a Bayesian Self-Attention Mechanism, which models epistemic uncertainty in three stages: patch embedding, uncertainty-aware self-attention computation, and adaptive regularization. This design significantly enhances the credibility of the model. Extensive experiments on the fuel cell datasets demonstrate the proposed method's outstanding performance. It achieves an average reduction of 36.31 % in root mean square error and an average compression of 83.39 % in the 95 % confidence interval, significantly outperforming existing methods. This approach offers a trustworthy basis for predictive maintenance in PEMFC systems, promoting a shift from “experience-based maintenance” to “reliable prognostics” in hydrogen energy applications.}, archive = {J_EAAI}, author = {Mengyu Liu and Zhe Cheng and Yu Yang and Niaoqing Hu and Guoji Shen and Yi Yang}, doi = {10.1016/j.engappai.2025.112444}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112444}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A reliable degradation prediction method for proton exchange membrane fuel cells based on uncertainty bayesian self-attention}, volume = {162}, year = {2025}, }
• (2025). Knowledge-data synergy enabling zero-shot composite fault diagnosis in sucker-rod pumping systems. EAAI, 162, 112443. (www)

  @article{ , abstract = {Zero-Shot Learning Accurate and comprehensive fault diagnosis is critical to ensuring production efficiency and extending equipment lifespan in oilfield operations. However, most existing diagnostic methods for sucker rod pumping systems are limited to single-fault identification, while the scarcity and high acquisition cost of compound fault samples hinder the application of supervised learning. To address these challenges, this paper proposes a zero-shot compound fault diagnosis framework that integrates domain knowledge with data-driven representations, enabling the identification of unseen fault using only single-fault samples. First, a unified semantic space is constructed by jointly embedding text-based fault descriptions and capsule-encoded load features. Then, a semantics-guided weakly supervised attribute composition strategy is introduced to enhance the completeness and discriminability of the semantic space. Finally, a bidirectional contrastive learning mechanism is established between visual encodings and semantic representations, and an adaptive multi-task loss weighting strategy is employed to optimize the overall framework efficiently. Experiments on real-world oilfield data demonstrate that the proposed method achieves an F1-score of 86.15 % under the zero-shot setting and 69.08 % under the generalized zero-shot setting, offering an effective and scalable solution for compound fault diagnosis in oil wells.}, archive = {J_EAAI}, author = {Xin-yan Wang and Li-ming Zhang and Kai Zhang and Cheng Cheng}, doi = {10.1016/j.engappai.2025.112443}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112443}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Knowledge-data synergy enabling zero-shot composite fault diagnosis in sucker-rod pumping systems}, volume = {162}, year = {2025}, }
• (2025). Edge computing and server-based high-precision flood level classification system. EAAI, 162, 112442. (www)

  @article{ , abstract = {Urban flooding and the resulting road water accumulation have become a significant threat to public transportation safety and the stability of municipal infrastructure. Traditional monitoring networks based on physical water level sensors suffer from low deployment density, high maintenance costs, and lagging response times. To address these shortcomings of traditional water accumulation monitoring systems, this study proposes an edge-computing intelligent monitoring system based on collaborative inference between the edge end (You Only Look Once version 5, YOLOv5) and the server end (Transform Vision Detection, TrVDet). A dual-modal perception architecture of “edge-end triggering and server-end precise analysis” has been constructed. At the edge end, the YOLOv5 model is deployed on embedded devices to achieve efficient preliminary screening of water accumulation, reducing dependence on the central server, lowering latency, and enhancing real-time response capabilities. On the server end, multi-object segmentation is performed on the detected water accumulation images, including roads, cars, motorcycles, and bicycles. Finally, a series of logical judgments is applied to determine the water accumulation level based on reference objects within the water. Since there is no publicly available dataset for target object recognition in flooded areas, we employed professional annotators to perform pixel-level labeling on the collected and organized flood data and constructed a multi-class target flood dataset (City Flood Segmentation, CityFloodSeg). Given the scarcity of moderate and severe water accumulation samples, we optimized the instance segmentation model TrVDet under the (A Visual Representation for Neon Genesis, EVA-02) framework and applied five data augmentation methods, including Mosaic and Flip, to expand the diversity of the dataset. Moreover, based on domain expert standards, we designed a logical judgment rule algorithm for model inference of water accumulation levels to classify the levels of water accumulation. Experimental results show that the server-end processing delay is stable within 0.4 s, capable of accurately judging different water accumulation risk levels. This provides centimeter-level real-time situational awareness for urban flood control decision-making and promotes the development of intelligent municipal infrastructure towards higher reliability and universality.}, archive = {J_EAAI}, author = {Ankang Lu and Runlong Cao and Yuanbin Wang and Wenjun Hu and Yuncan Gao and Zhifeng Hu and Ying Zang}, doi = {10.1016/j.engappai.2025.112442}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112442}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Edge computing and server-based high-precision flood level classification system}, volume = {162}, year = {2025}, }
• (2025). WeedNet-X: A lightweight field weed detection algorithm. EAAI, 162, 112441. (www)

  @article{ , abstract = {Accurately distinguishing field weeds from crops and locating weed positions are critical prerequisites in automated weed control operations. However, weed detection and localization in unstructured field environments with complex lighting remains challenging. Firstly, data-driven deep learning based algorithms usually have a high dependence on a large number of training samples, and there are huge differences between field weeds and crops in different regions, growth cycles, and types. In addition, the conflict between hardware performance and computation cost makes it difficult for existing weed detection algorithms to maintain both detection accuracy and speed on low-performance platforms. All these problems increase the difficulty of detection. To solve the above problems, we first construct a medium-to-large weed dataset using an open-source agricultural image dataset and collect field data. Subsequently, we have proposed a lightweight weed detection algorithm using the ShuffleNetv2 network as the backbone network, with a multi-scale pyramid network, and the overall network algorithm is named WeedNet-X. The number of model parameters and the computational volume of the algorithm are only 0.57 million and 0.48 Giga floating point operations (GFLOPs), respectively. On the two constructed datasets, the mean Average Precision (mAP) of the algorithm can reach 86.31 % and 80.98 %, respectively, which are improved by 0.61 % and 3.10 % compared to the baseline model. Finally, the hardware and software systems for weed detection verify the excellence of the proposed algorithm in terms of practical performance.}, archive = {J_EAAI}, author = {Yong Li and Ao Ke and Zhiqiang Guo and Qingji Tan and Jingchao Yang}, doi = {10.1016/j.engappai.2025.112441}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112441}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {WeedNet-X: A lightweight field weed detection algorithm}, volume = {162}, year = {2025}, }
• (2025). Data-driven robust intertwined wheat supply chain: Redesigning a viable network for long-term geopolitical conflicts. EAAI, 162, 112440. (www)

  @article{ , abstract = {The recent human-made conflict in 2022 severely damaged Ukraine's infrastructure, causing significant instability in the food supply chain. This crisis was further exacerbated by trade bans imposed on another major global wheat exporter. Since wheat production and export are intrinsically linked, particularly in times of crisis, it is essential to adopt the concept of an intertwined supply chain. Accordingly, this study proposes an intertwined supply chain framework for the production and export of wheat during long-term disruptions. To enhance the viability of this intertwined system, the study introduces three key strategies. First, it addresses long-term disruptions and operational risks by employing redundancies and data-driven robust optimization techniques, where uncertainty sets are generated using a support vector clustering model. Second, the proposed supply chain accounts for freshwater resource limitations by integrating water resilience measures. Third, as the framework operates within a global context, it incorporates a comprehensive model that considers exchange rates, taxation, foreign demand points, and international trade responsibilities. To optimize these strategies, two multi-objective optimization models are developed and solved using an epsilon-constraint method. A cardinality-based measure is introduced to efficiently represent the Pareto front, offering decision-makers valuable insights into non-dominated solutions. The results are divided into analyses of wheat production, export, and their combined network. Individual analyses assess network setup, viability, and uncertainty control, while the integrated analysis examines sensitivity and interdependence. Overall, improving water use, managing risks, and designing a resilient, interconnected system can greatly strengthen the wheat supply chain during long-term crises.}, archive = {J_EAAI}, author = {Hani Gilani and Mehrdad Mohammadi and Tom Van Woensel and Hadi Sahebi}, doi = {10.1016/j.engappai.2025.112440}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112440}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Data-driven robust intertwined wheat supply chain: Redesigning a viable network for long-term geopolitical conflicts}, volume = {162}, year = {2025}, }
• (2025). A multitask learning-based fault detection approach for high impedance fault in resonant distribution networks. EAAI, 162, 112439. (www)

  @article{ , abstract = {Addressing high impedance faults (HIF) in resonant distribution networks remains a formidable challenge. This paper introduces a multitask learning-based approach integrated with a multitask fault detection network (MTFD-Net), employing three task-specific heads—classification, segmentation, and regression—to enable precise fault detection. MTFD-Net utilizes zero-sequence voltage data and a sliding time window to perform initial coarse classification, which allows the classification head to determine whether a permanent HIF has occurred. Upon detection, MTFD-Net proceeds to pinpoint potential fault moments through the outputs of the segmentation head. The regression head further refines these moments by predicting a reference moment and calculating the distance to each potential fault, effectively isolating the exact fault moment. An industrial prototype was developed and rigorously tested on a 10 kV system, where MTFD-Net demonstrated superior performance, achieving an accuracy of 0.976, an intersection over union of 0.984, and an absolute detection deviation of 5.20 ms. Operating efficiently with inference times ranging from 9.69 to 15.45 miliseconds on a Raspberry Pi 4B, MTFD-Net surpasses existing methods in accuracy, F1-score, sensitivity, specificity, and detection accuracy, providing a robust solution for HIF detection in resonant distribution networks.}, archive = {J_EAAI}, author = {Jian-Hong Gao and Mou-Fa Guo and Shuyue Lin and Duan-Yu Chen}, doi = {10.1016/j.engappai.2025.112439}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112439}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multitask learning-based fault detection approach for high impedance fault in resonant distribution networks}, volume = {162}, year = {2025}, }
• (2025). Carbon trading price prediction with spikes: A novel hybrid model framework using heuristic multi-head attention convolutional bidirectional recurrent neural network. EAAI, 162, 112438. (www)

  @article{ , abstract = {Accurate forecasting of carbon trading prices (CTPs) with spikes is crucial for developing carbon emission reduction policies and planning corporate investments. However, most existing CTP approaches usually focus on designing a cutting-edge model without considering spike prediction. Therefore, this paper presents a novel heuristic optimization-based hybrid model framework for CTP prediction with spikes. First, random forest is exploited to identify the relevant features of spikes and non-spikes for CTPs, and categorical boosting is employed to predict the spike occurrences of CTPs. Then, a novel hybrid model based on multiple linear regression, categorical boosting, and two dimensions convolutional neural network and bidirectional gated recurrent unit with multi-head regularized attention mechanism (2DCNN-BiGRU-MRA) is proposed to predict spikes and non-spikes for CTPs. In this model, multiple linear regression and categorical boosting are respectively applied to capture the linear and complex nonlinear features of the CTPs, in which their prediction results and deviations are integrated into the 2DCNN-BiGRU-MRA model as relevant features. The proposed 2DCNN-BiGRU-MRA can learn the spatiotemporal features and enhance representation capabilities by introducing 2DCNN, BiGRU, and MRA, thereby improving the accuracy of CTP prediction. In addition, to construct appropriate model hyperparameters of 2DCNN-BiGRU-MRA, the strength honey badger algorithm based on the adaptive momentum estimation is proposed to optimize the hyperparameters of 2DCNN-BiGRU-MRA. Finally, the proposed framework is tested on the actual data of European Union emissions trading and the carbon market in Hubei, China, and case studies have confirmed the superiority and achievable local interpretability of the proposed hybrid model framework.}, archive = {J_EAAI}, author = {Rongquan Zhang and Siqi Bu and Gangqiang Li and Min Zhou}, doi = {10.1016/j.engappai.2025.112438}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112438}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Carbon trading price prediction with spikes: A novel hybrid model framework using heuristic multi-head attention convolutional bidirectional recurrent neural network}, volume = {162}, year = {2025}, }
• (2025). Inter-graph and intra-graph: Utilizing global financial markets and constituent stocks for stock index prediction. EAAI, 162, 112437. (www)

  @article{ , abstract = {Stock index prediction is a significant yet difficult undertaking due to its incorporation of complex and diverse information. Following the implementation of Graph Neural Networks in financial data analysis, numerous researchers have focused on the node-level task of forecasting individual stock movements by analyzing the relationships between stocks. However, two key challenges remain: first, realizing different speeds of feature propagation among nodes in graph representation learning; second, predicting stock indices by extracting and aggregating fluctuations from constituent stocks through graph-level tasks remains unaddressed. To tackle these challenges, this paper proposes a novel spatio-temporal prediction framework combining both node-level and graph-level tasks. The framework includes two types of graphs: inter-graph and intra-graph, which combine information from the micro, meso, and macro dimensions. For the inter-graph at the node level, we introduce the Granger causality test as an innovative node filtering method, which realizes the propagation of features between nodes with different strengths and speeds in the process of graph representation learning. For the intra-graph at the graph level, we examine various graph pooling methods and pooling proportions of stock index constituents to enhance the interpretability of the results and to provide new theoretical insights for stock index prediction. In conclusion, we develop the Graph Representation Learning-based Long Short-Term Memory (GRL-LSTM) model for forecasting stock index movements, and demonstrate the superiority of our approach on four major Chinese stock markets.}, archive = {J_EAAI}, author = {Yong Shi and Yunong Wang and Jie Wu}, doi = {10.1016/j.engappai.2025.112437}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112437}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Inter-graph and intra-graph: Utilizing global financial markets and constituent stocks for stock index prediction}, volume = {162}, year = {2025}, }
• (2025). A novel data-efficient double deep Q-network framework for intelligent financial portfolio management. EAAI, 162, 112436. (www)

  @article{ , abstract = {Navigating the complexities of dynamic and uncertain financial markets demands intelligent systems capable of learning profitable strategies amidst risk and volatility. While Deep Q-Networks (DQN) offer a foundation for such systems, they often suffer from overestimation bias, training instability, and poor generalization in noisy financial environments. To address these challenges, this work introduces Portfolio Double Deep Q-Network (PDQN), a novel architecture inspired by recent advancements in reinforcement learning. PDQN enhances portfolio management by integrating Double Q-Learning to reduce overestimation, alongside Leaky ReLU activation, Xavier initialization, Huber loss, and dropout regularization to improve learning stability and generalization. Unlike prior methods that rely on large datasets and heavy computational infrastructure, PDQN achieves competitive—and often superior—performance using substantially less training data and lightweight infrastructure, making it well-suited for real-world, resource-constrained financial applications. Distinct from conventional approaches, PDQN uses separate networks to adapt portfolio decisions across varying market conditions. Empirical results across multiple market years show that PDQN often outperforms baseline strategies, including classic DQN and Buy-and-Hold, across key metrics such as Sharpe ratio, Sterling ratio, and cumulative return. PDQN—like all data-driven models—exhibits room for improvement under highly irregular or extreme financial scenarios. These observations suggest promising directions for future refinement and increased robustness, without detracting from the model's practical effectiveness and competitive edge.}, archive = {J_EAAI}, author = {Mahshad Alidousti and Morteza Khakzar Bafruei and Amir Hosein Afshar Sedigh}, doi = {10.1016/j.engappai.2025.112436}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112436}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel data-efficient double deep Q-network framework for intelligent financial portfolio management}, volume = {162}, year = {2025}, }
• (2025). Unlocking few-shot wind speed prediction through a novel end-to-end transfer learning paradigm based on decomposition and gating information fusion. EAAI, 162, 112435. (www)

  @article{ , abstract = {Accurate wind speed prediction is one of the key technologies for achieving intelligent and sustainable development in the engineering field. In the field of wind speed prediction, we are confronted with a challenging few-shot prediction problem. Specifically, due to the fact that some wind turbines in wind farms are newly established or there is data loss during the data collection process, these turbines only contain a small amount of wind speed data. This scarcity of data poses great difficulties for the prediction work, and traditional prediction methods often fail to achieve the desired prediction accuracy. In order to overcome the above difficulties, we propose an novel prediction paradigm of end-to-end transfer learning based on data decomposition and gated information fusion. We use the Fourier transform to find the source domain similar to the target domain to achieve feature alignment. Then, we pre-train the model on the source domain and transfer this model to the target domain, thus solving the problem of low prediction accuracy when directly predicting the target domain. In the first step, the data is decomposed and denoised by using the Variational Mode Decomposition. According to the sample entropy, the decomposed data is reorganized into three frequency components. Each component is input as an independent channel into the end-to-end prediction model. Firstly, the features of each channel are expanded to a high-dimensional space through the Multilayer Perceptron. Then, the gating mechanism is utilized to mix the features of the three channels into the features of one channel, thus achieving information fusion. Finally, the prediction result of the end-to-end model is output through the Gated Recurrent Unit. In the second step, the model pre-trained on the source domain is transferred to the small-sample target domain. The Dynamic Time Warping and cosine similarity are used to quantify the similarity of each channel between the two domains. The parameters of the channels with high similarity are locked, and at the same time, the parameters of other channels are fine-tuned to output the final prediction result. In addition, multiple sets of comparative experiments conducted using the wind speed data from wind farms in Queensland, Australia, have demonstrated the superiority of this prediction paradigm. Our strategy outperforms various baseline models in all three sets of data. Moreover, ablation experiments have proven the effectiveness of each component in this framework in improving prediction accuracy, opening up a new path for solving the difficult problem of few-shot prediction in practical engineering.}, archive = {J_EAAI}, author = {Xiaoyue Dong and Zhirui Tian}, doi = {10.1016/j.engappai.2025.112435}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112435}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unlocking few-shot wind speed prediction through a novel end-to-end transfer learning paradigm based on decomposition and gating information fusion}, volume = {162}, year = {2025}, }
• (2025). A hyperparameter-fusion neural networks for deposition prediction. EAAI, 162, 112434. (www)

  @article{ , abstract = {As integrated circuit manufacturing processes develop into the nanometer scale, precise control and prediction of the deposition process have become crucial. Nanoscale manufacturing imposes unprecedentedly high demands on film quality, uniformity, and consistency, presenting significant challenges to traditional control and prediction methodologies. This study proposes a novel approach that, for the first time, formulates the thin-film deposition process as a video prediction task, enabling the use of deep learning for morphological forecasting under varying process conditions, and introduces a novel hyperparameter-fusion neural network, referred to as DepositionNet (DepoNet). Unlike conventional video prediction models, DepoNet specifically accounts for the influence of deposition parameters on the entire simulation process. We have incorporated a novel Hyper Projector that allows the model to flexibly adapt to varying deposition conditions and material characteristics. Through comprehensive comparative experimental analyses, we demonstrate that DepoNet significantly outperforms existing deep-learning models and achieves a mean squared error of 17.34, representing a 3.67% improvement over the second best model and a 1,435 × speedup over physics-based methods, thereby validating its exceptional generalization capability. Extensive experiments reveal that the model maintains high performance even under conditions of limited training data, for instance, achieving a peak signal-to-noise ratio (PSNR) of 41.516 decibels (dB) when trained with only 20% of the available data.}, archive = {J_EAAI}, author = {Li Ding and Kun Pang and Junjie Li and Hua Shao and Nan Liu and Rui Chen and Zhiqiang Li and Zhenjie Yao and Ling Li}, doi = {10.1016/j.engappai.2025.112434}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112434}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A hyperparameter-fusion neural networks for deposition prediction}, volume = {162}, year = {2025}, }
• (2025). Overcoming challenges in leveraging blockchain technology: Entropy-based q-rung orthopair fuzzy model for benchmarking application barriers. EAAI, 162, 112433. (www)

  @article{ , abstract = {Blockchain technology has emerged as a transformative solution across industries, delivering enhanced transparency, security, and operational efficiency. Nevertheless, its adoption remains hindered by significant challenges, especially in complex, data-intensive domains such as logistics. This study introduces a novel integration of the entropy-based q-rung orthopair fuzzy compromise ranking of alternatives from distance to ideal solution (CRADIS) approach to systematically evaluate and prioritize key barriers to blockchain adoption. The innovation of this work lies in applying q-rung orthopair fuzzy sets which are particularly capable of handling higher degrees of uncertainty and hesitancy, and then integrated with entropy for objective criterion weighting and CRADIS for robust decision-making. A real-world case study is presented, involving five critical barriers, lack of legal and regulatory frameworks, high implementation costs, technological scalability issues, data privacy and security concerns, and cultural resistance to change evaluated against eight decision criteria. The entropy weighting revealed regulatory clarity (0.168) and security (0.154) as the most influential factors, while the CRADIS ranking identified a lack of legal frameworks as the top barrier. This framework provides a transparent, data-driven method for decision-makers to identify and prioritize adoption challenges, particularly in uncertain and multi-faceted environments. By demonstrating the model’s applicability and precision, the study contributes to the emerging body of literature on blockchain integration and supports organizations in navigating the transition towards decentralized technologies.}, archive = {J_EAAI}, author = {Sana Shahab and Naoufel Kraiem and Ashit Kumar Dutta and Mohd Anjum and Vladimir Simic and Dragan Pamucar}, doi = {10.1016/j.engappai.2025.112433}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112433}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Overcoming challenges in leveraging blockchain technology: Entropy-based q-rung orthopair fuzzy model for benchmarking application barriers}, volume = {162}, year = {2025}, }
• (2025). Dual angle magnetic arc blow estimation in keyhole tungsten inert gas welding using high dynamic range imaging and a lightweight vision transformer network with coordinate attention and multiple auxiliary branches. EAAI, 162, 112432. (www)

  @article{ , abstract = {In high current Keyhole Tungsten Inert Gas (K-TIG) welding, magnetic arc blow frequently causes severe defects such as lack of fusion and undercut, which seriously affect weld formation quality. Conventional visual sensing systems are limited by dynamic range, making it difficult to capture arc morphology, while single angle descriptors fail to represent nonlinear deflection and lightweight convolutional models struggle with long range dependencies. To address these challenges, this study employs a High Dynamic Range (HDR, 120 decibel [dB]) imaging system to capture detailed arc variations and proposes a lightweight Vision Transformer (ViT) network with embedded Coordinate Attention (CA) and multiple auxiliary branches for real time angle estimation. A custom magnetic excitation system enables controllable arc blow simulation and consistent data acquisition. The method introduces a dual angle representation, namely the maximum curvature angle ( θ curv ) and the equivalent deviation angle ( θ eq ), to comprehensively describe arc geometry. The Artificial Intelligence (AI) framework integrates segmentation, keypoint localization, and regression tasks to improve accuracy and robustness. Trained on a self constructed HDR dataset containing 3,191 annotated images, model achieves a mean absolute error (MAE) of 1 . 12 ° , a root mean square error (RMSE) of 2 . 84 ° , a determination coefficient ( R 2 ) of 0.96, and a per frame inference latency of 12.96 ms (ms) on an NVIDIA RTX 2080Ti graphics processing unit (GPU). These results demonstrate that AI based methods combined with HDR imaging cannot only achieve accurate monitoring of welding arc states, but also provide potential support for closed loop control in all position welding applications.}, archive = {J_EAAI}, author = {Xiyin Chen and Xiaohu Zhang and Yonghua Shi and Yuxiang Huang and Junjie Pang}, doi = {10.1016/j.engappai.2025.112432}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112432}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dual angle magnetic arc blow estimation in keyhole tungsten inert gas welding using high dynamic range imaging and a lightweight vision transformer network with coordinate attention and multiple auxiliary branches}, volume = {162}, year = {2025}, }
• (2025). Plug-and-play dynamic optimization for three-dimensional gaussian generation. EAAI, 162, 112431. (www)

  @article{ , abstract = {Recent advancements in Three-Dimensional (3D) asset generation have demonstrated remarkable progress in generation efficiency, enabling transformative applications across creative industries and mission-critical domains including autonomous systems. Current 3D asset generation primarily employs Score Distillation Sampling (SDS) to derive 3D priors from Two-Dimensional (2D) diffusion models. While this contribution ensures high generation quality, it is time-consuming. Recent methods have utilized 3D Gaussian Splatting for image rendering, which, despite enhancing generation speed, compromised on quality. Our method aims to balance the quality and speed of 3D asset generation by designing a plug-and-play optimization process that combines the strengths of both methods. We propose a rapid 3D Gaussian generation framework that begins with constructing a pipeline to generate multi-view images from text input using pre-trained generative models. Then our method utilizes 3D Gaussian Splatting for quick 3D asset initialization and subsequently performs detail optimization using Gaussian Filter and SDS-based 2D diffusion model optimizer. Additionally, we have optimized the loss function for 3D Gaussian Splatting and ensured the entire optimization process is plug-and-play, offering high generation quality and speed. Our method demonstrates strong adaptability in representative single-object 3D Gaussian generation tasks, indicating promising generalization potential. Achieving high-quality 3D generation on a single Graphics Processing Unit (GPU), our framework outperforms most popular optimization-based models in generation speed (5 × speedup +). Furthermore, when juxtaposed with the latest inference-based models, our optimization architecture offers a notable enhancement in generation quality (Contrastive Language-Image Pre-Training Score 33.8 vs. 27.3) within an acceptable amount of time.}, archive = {J_EAAI}, author = {Qixuan Li and Haoyang Li and Chao Wang and Yang Zhou and Yan Peng}, doi = {10.1016/j.engappai.2025.112431}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112431}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Plug-and-play dynamic optimization for three-dimensional gaussian generation}, volume = {162}, year = {2025}, }
• (2025). A diffusion model using semantic and sketch information for anomaly detection. EAAI, 162, 112430. (www)

  @article{ , abstract = {In anomaly detection, methods that employ diffusion models for anomaly localization and reconstruction have demonstrated significant achievements. However, these methods face challenges such as the misclassification of multiple types of anomalies and the inability to effectively reconstruct large-scale anomalies due to the absence of semantic and sketch information from the original images. To tackle these challenges, we propose a framework, A Diffusion Model using Semantic and Sketch Information for Anomaly Detection (DSAD), which includes a semantic and sketch-guided network (SSG), a pre-trained autoencoder, and Stable Diffusion (SD). Initially, within SSG, we introduce a Semantic & Sketch Feature Fusion Module to enhance the model’s comprehension of the original images and present a Multi-scale Feature Fusion Module to maximize reconstruction accuracy. Subsequently, we connect SSG with the denoising network in SD in order to guide the network in reconstructing anomalous regions. Experiments on MVTec-AD dataset demonstrate the effectiveness of our approach which surpasses the state-of-the-art methods. The dataset and code are available at https://github.com/QinLi-STUDY/DSAD/tree/master .}, archive = {J_EAAI}, author = {Li Qin and Zhenyu Yin and Feiqing Zhang and Chunhe Song and Xiaoqiang Shi}, doi = {10.1016/j.engappai.2025.112430}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112430}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A diffusion model using semantic and sketch information for anomaly detection}, volume = {162}, year = {2025}, }
• (2025). A real-time spatiotemporal error compensation framework for face gear grinding. EAAI, 162, 112429. (www)

  @article{ , abstract = {Geometric and thermal errors critically affect the precision of face gear grinding, yet current modeling approaches are computationally intensive and lack real-time adaptability. This study proposes a real-time spatiotemporal error compensation framework for face gear grinding. A closed-loop feedback mechanism is introduced to adaptively update compensation intensity based on residual error feedback, ensuring robustness and efficiency under fluctuating machining conditions. Moreover, a novel spatial-temporal thermal error model is developed by integrating Taylor-graph convolutional network and modified-long short term memory network to capture both node-level spatial fusion and long-term temporal dependencies. High-order terms in geometric error modeling are eliminated using a vector decomposition and truncation-based approach, significantly reducing computational complexity. Furthermore, a high-efficiency multi-source error-tooth flank mapping model is developed based on vector decomposition and truncation function methods, enabling accurate prediction with reduced computational cost. To identify dominant error contributors, an improved Morris-based sensitivity analysis method is integrated, distinguishing geometric and thermal errors affecting tooth flank deviation. Experimental results demonstrate sub-65 ms real-time response, 24.2 μm maximum error reduction, and robust adaptability under fluctuating machining conditions. Compared with recent gear-flank compensation studies, the proposed closed-loop framework achieves a 63.4 % reduction in maximum normal flank error under real machining and <65 ms response latency. This level is comparable to reported reductions based on grid-aggregated metrics in spiral bevel gears (76.82 % reduction of the sum of absolute grid errors), while additionally ensuring real-time, delay-aware execution. These findings validate the proposed system's potential for precision, real-time compensation in multi-axis manufacturing environments.}, archive = {J_EAAI}, author = {Jialan Liu and Chi Ma and Mingming Li and Jialong He and Giovanni Totis and Chunlei Hua and Gangwei Cui and Liang Wang and Ruijun Xue and Zhi Tan and Jun Yang and Kuo Liu and Yuansheng Zhou and Jianqiang Zhou and Xiaolei Deng and Shengbin Weng}, doi = {10.1016/j.engappai.2025.112429}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112429}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A real-time spatiotemporal error compensation framework for face gear grinding}, volume = {162}, year = {2025}, }
• (2025). Multilayer perceptron-based offspring prediction model for constrained multi-objective optimization. EAAI, 162, 112428. (www)

  @article{ , abstract = {Constrained multi-objective optimization problems generally have both multiple constraint violations and conflicting objective functions. Some of them not only have sparse feasible regions, but also are difficult to converge. For these problems, the evolutionary operators used in traditional constrained multi-objective evolutionary algorithms (CMOEAs) are difficult to generate solutions with ideal quality. Therefore, this paper proposes a multilayer perceptron-based offspring prediction model for constrained multi-objective optimization (MOPCMO). Specifically, an evolutionary direction guidance strategy is designed that utilizes historical populations as training data to train a multilayer perceptron, which guides the evolution of the population by predicting and generating offspring, thereby improving the overall evolutionary efficiency of the algorithm. In addition, as the population iterates, evolutionary direction guidance strategy adaptively transforms the training data of multilayer perceptron. Finally, the multilayer perceptron is intermittently updated and uses an evolutionary direction guidance strategy to generate promising offspring, guiding the algorithm to achieve efficient search. Compared with seven state-of-the-art CMOEAs on 33 benchmark test problems and 8 engineering application problems, MOPCMO achieves excellent performance.}, archive = {J_EAAI}, author = {Qianlong Dang and Ruihuan Luo and Linlin Xie and Xiaochuan Gao and Weiting Bai}, doi = {10.1016/j.engappai.2025.112428}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112428}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multilayer perceptron-based offspring prediction model for constrained multi-objective optimization}, volume = {162}, year = {2025}, }
• (2025). Fault diagnosis via multi-sensor fusion with auxiliary contrastive learning and phased fine-tuning. EAAI, 162, 112427. (www)

  @article{ , abstract = {Typically, deep learning-based fault diagnosis models fail to fully utilize the potential information in large amounts of normal state data and encounter difficulties when learning from limited fault samples. To address these challenges, this study proposes an auxiliary contrastive learning framework designed for multi-sensor data. The framework incorporates auxiliary classifiers after each sensor-specific branch to enhance feature representation, and enables model pretraining using only normal condition data. In addition, a phased fine-tuning strategy is developed, which combines full-model fine-tuning with lightweight adapter tuning to improve the adaptability of the fine-tuning process. A novel multi-sensor data augmentation technique is also introduced to enrich the contrastive learning tasks by generating structurally diverse negative samples. By enabling the effective utilization of normal condition data in model training, the proposed framework offers a new perspective for fault diagnosis applications. Experimental results on three benchmark datasets demonstrate that the proposed method significantly improves the generalization capability of the pre-trained model. Furthermore, the phased fine-tuning strategy exhibits high adaptability to the target tasks. Compared to other data fusion methods, the proposed auxiliary contrastive learning framework achieves notable performance advantages.}, archive = {J_EAAI}, author = {Yulin Jin and Xiaochuan Luo and Xiangwei Kong and Yulin Zhang}, doi = {10.1016/j.engappai.2025.112427}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112427}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fault diagnosis via multi-sensor fusion with auxiliary contrastive learning and phased fine-tuning}, volume = {162}, year = {2025}, }
• (2025). Damage and energy dissipation prediction and multi-objective optimization design of latticed concrete-filled steel tube column-composite box girder joints based on extreme gradient boosting. EAAI, 162, 112426. (www)

  @article{ , abstract = {To optimize joint performance, a finite element (FE) model is developed based on low-cycle reciprocating load tests of latticed concrete-filled steel tubular (CFST) column-composite box girder joints. The FE-predicted hysteresis curves are compared with test results to verify model accuracy, and a data set is established accordingly. Extreme Gradient Boosting (XGBoost) algorithm is used for training and prediction, and compared with the traditional machine learning (ML) algorithm, the superiority of the XGBoost algorithm is manifested. The XGBoost algorithm is then used to predict the damage and energy values of the joint under more different parameter combinations, with the largest ratio of damage value to energy dissipation value selected as the optimal combination of the joints within the variation range of the six parameters. The results show that the FE model correlates well with the test results and can therefore be used to generate a data set. The prediction accuracy of XGBoost algorithm has high accuracy of more than 99 % in predicting damage and energy dissipation values and can thus be used for joint prediction research. Compared with other ML algorithms, XGBoost has the best prediction performance and superiority. Within the variation range of the six parameters, the ratio of damage value to energy dissipation value is the largest when the concrete strength, longitudinal bar diameter, concrete slab thickness, box girder strength, axial compression ratio, and transverse stiffener strength are respectively 30 Mega Pascal (MPa), 12 mm (mm), 90 mm, 390 MPa, 0.3, and 455 MPa.}, archive = {J_EAAI}, author = {Zhi Huang and Xin Deng and Juan Chen and Xiang Li and Lizhong Jiang and Yohchia Frank Chen and Yuner Huang and Lin Chen}, doi = {10.1016/j.engappai.2025.112426}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112426}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Damage and energy dissipation prediction and multi-objective optimization design of latticed concrete-filled steel tube column-composite box girder joints based on extreme gradient boosting}, volume = {162}, year = {2025}, }
• (2025). Multi-dimensional feature fusion network design and performance optimisation for small target detection. EAAI, 162, 112425. (www)

  @article{ , abstract = {Due to the long distance of image acquisition, high imaging resolution, complex feature background, shooting angle, etc. The result is that there are few features available for small targets and they are easily interfered by background noise, which poses a challenge to the detection of small targets. To address the above problems, this paper proposes a target detection network (Convolution-based Small Target Detection Network, CSTDNet) with enhanced feature information, which integrates a multi-dimensional information fusion strategy for small target features. An all-round efficient feature fusion mudule (AeFusion) is introduced, which emphasises the fusion of multi-dimensional feature information, enhances the model's ability to focus on key information and suppress redundant information, and strengthens the ability to characterise local features and details, improving the effectiveness of the information and computational efficiency. In order to further enhance the location-awareness capability in cross-layer interaction, this paper introduces a novel decoupling head (Self-aware task decomposition for fine-grained feature sharing, STFS), which improves the accuracy of the small-target classification and localisation tasks through efficient detail sharing and task auto-alignment functions. And localisation tasks through efficient detail sharing and task auto-alignment. This study evaluates the effectiveness of the algorithm on five different scenarios containing small target datasets. Experimental results show that CSTDNet achieved improvements of 6.6 %, 5.8 %, 5.8 %, 5.5 %, and 5.6 % over the baseline model in terms of the mean average precision (mAP@0.5) metric on the Visdrone 2019, BDD100K, WiderPerson, SODA10M, and AppleDatas datasets, respectively, demonstrating stronger detection performance.}, archive = {J_EAAI}, author = {Xiaoyao Yang and Wenyang Zhao and Pengchao Sun and Wenda Zhao and Wenlong Yang}, doi = {10.1016/j.engappai.2025.112425}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112425}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-dimensional feature fusion network design and performance optimisation for small target detection}, volume = {162}, year = {2025}, }
• (2025). FDFNet: Feature-decision dual fusion network for intelligent fault diagnosis of rotating machinery under varying speed conditions. EAAI, 162, 112423. (www)

  @article{ , abstract = {Fault diagnosis is a critical technique for enhancing the reliability and security of rotating machinery. Existing diagnosis methods still have restricted generalization performance under speed variations owing to inadequate utilization of multisensor information. To address this problem, a novel feature-decision dual fusion network is proposed for fault diagnosis of rotating machinery under varying speed conditions. First, for each sensor, the frequency information learner is built to simultaneously extract global and local frequency domain features using global and local feature encoders. These features are then fused through a cross-attention mechanism to generate a sensor-specific initial classification decision. Subsequently, these individual sensor-wise decisions are fed into the decision dynamic ensemble to yield final fault diagnosis result. Moreover, an adaptive optimization strategy is designed to guide model learning generalizable features by flexibly adjusting the sensor loss weights during training. Finally, the effectiveness of the proposed method is validated on bearing and gearbox datasets. Experimental results demonstrate that the proposed method exhibits superior generalization and robustness for fault diagnosis under varying speed conditions.}, archive = {J_EAAI}, author = {Qi Deng and Zuoxiu Zhang and Xuyuan Tu and Zimuzhi Wang and Jun Wu}, doi = {10.1016/j.engappai.2025.112423}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112423}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {FDFNet: Feature-decision dual fusion network for intelligent fault diagnosis of rotating machinery under varying speed conditions}, volume = {162}, year = {2025}, }
• (2025). Resource-efficient cross-subject emotion recognition from electroencephalogram via spiking domain discriminators. EAAI, 162, 112422. (www)

  @article{ , abstract = {Real-time emotion recognition poses a significant challenge in electroencephalogram (EEG) based emotion recognition, as it requires the immediate processing of EEG data. This necessity imposes substantial demands on the model’s resource consumption. To address this issue, this paper introduces a novel approach to EEG emotion recognition using a Cross Domain Spiking Convolutional Network (CDSCN), focusing on developments in the design of the spiking convolutional block. To address individual differences, the CDSCN incorporates Z-Score normalization at the feature level and introduces a spiking domain discriminator at the model level. These innovations aim to mitigate variations in data distribution across individuals and domains, thereby enhancing the model’s robustness and generalizability. Additionally, the CDSCN introduces a novel pooling fusion layer within the spiking convolutional block to optimize computational efficiency while preserving discriminative performance. Experimental evaluations on two publicly available datasets validate the effectiveness of the proposed CDSCN in achieving both accurate and generalized emotion recognition.}, archive = {J_EAAI}, author = {Dongdong Li and Shengyao Huang and Yujun Shen and Zhe Wang}, doi = {10.1016/j.engappai.2025.112422}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112422}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Resource-efficient cross-subject emotion recognition from electroencephalogram via spiking domain discriminators}, volume = {162}, year = {2025}, }
• (2025). Lightweight deep learning based solar cells defect detection using electroluminescence images. EAAI, 162, 112421. (www)

  @article{ , abstract = {Solar cells are the fundamental core energy harvesting components in photovoltaic (PV) power generation stations. In view of the capability of detecting the invisible defects, the electroluminescence (EL) imaging is broadly used in the production lines of solar cells, based on which the deep learning technique is introduced to implement automatic defect detection and classification. However, the current deep learning models feature high complexity and require much computation resources, which are difficult to deploy in edge devices for real time applications. To tackle this issue, we proposed a novel lightweight and high-precision deep learning model named Cross Stage Partial Photovoltaic-You Only Look Once (CSPV-YOLO) based on the deep learning framework You Only Look Once v5 (YOLOv5) to enable the real-time solar cell defect detection. Firstly, a new module Cross Stage Partial C5 (CSPC5) is proposed to replace the initial C3 module in the YOLOv5 network to enhance the network accuracy in recognizing different types of defects. Secondly, a novel Spatial Pyramid Pooling with Cross Stage Partial (SPPFCSP) module is designed to replace the original Spatial Pyramid Pooling Fast (SPPF), which boosts the network feature extraction capabilities from defect targets at multiple scales and facilitates a more efficient integration of multiscale features. Finally, the original loss function of YOLOv5 is replaced by the Scylla intersection over union (SIoU) function to optimize the training model. The proposed models have been validated and intensively compared with many other state-of-the-art models on two public datasets. Firstly, results of experiments on the public Pascal Visual Object Classes (PASCAL VOC) 2007 datasets demonstrate that the proposed SPPFCSP block is obviously superior to other Spatial Pyramid Pooling blocks for the most state-of-the-art YOLO detectors, which can significantly improve the detection accuracy. The comparison results of experiments on the public Photovoltaic Electroluminescence Anomaly Detection Dataset (PVEL-AD) that includes 12-class defects obviously indicate that the proposed CSPV-YOLO model is better than many state-of-the-art models and achieves 91.5 % average precision (AP) and frames per second (FPS) of 177.8 on with only 2.2 million (M) parameters. Hence, it is suitable for the deployment on edge devices for real-time applications.}, archive = {J_EAAI}, author = {Zhicong Chen and Tianxiang Chen and Haoxin Zheng and Lijun Wu and Shuying Cheng and Peijie Lin}, doi = {10.1016/j.engappai.2025.112421}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112421}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Lightweight deep learning based solar cells defect detection using electroluminescence images}, volume = {162}, year = {2025}, }
• (2025). Multiple scattering-assisted data-driven full waveform inversion for pipeline blockage detection. EAAI, 162, 112420. (www)

  @article{ , abstract = {Blockages in urban water supply systems (UWSS) cause operational inefficiencies and require timely detection and intervention. Hydraulic transients offer an efficient method for diagnosing UWSS. However, existing transient-based approaches face limitations: they either rely on gradient-descent optimization algorithms, which are prone to convergence issues due to noise, or they assume weak blockage-induced scattering that only applies to small blockages, thereby restricting their diagnostic scope. This study introduces a novel approach that leverages multiple scattering of transient wavefields to enhance full waveform inversion (FWI) for accurate pipeline blockage detection and profiling. Recognizing that blockage characteristics – such as length, size, and location – significantly alter the system’s transient response, we perturb these parameters to generate a comprehensive dataset using a numerical model. This dataset is used to train a machine learning (ML) model designed to associate multiple scattered wavefield patterns with specific blockage characteristics. The ML model performs full waveform inversion to predict spatial variations in pipeline area, thereby revealing detailed blockage profiles. We validate the effectiveness of our method through numerical simulations and experimental data, which account for multiply scattered wavefields influenced by visco-elastic and frictional damping effects. Specifically, we quantify prediction accuracy by reporting errors in key blockage parameters such as length and radial extent, which directly influence pipeline performance. Additionally, we assess the framework’s resilience to noise, a common challenge in practical applications. The proposed framework demonstrates that ML can effectively interpret complex scattered wavefield data related to pipeline defects, paving the way for advancements in diagnosing large-scale water networks.}, archive = {J_EAAI}, author = {Utban Ahmed and Muhammad Waqar and Fedi Zouari and Liyou Luo and Moez Louati and Jensen Li and Mohamed S. Ghidaoui}, doi = {10.1016/j.engappai.2025.112420}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112420}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multiple scattering-assisted data-driven full waveform inversion for pipeline blockage detection}, volume = {162}, year = {2025}, }
• (2025). A streaming variable neural speech codec. EAAI, 162, 112418. (www)

  @article{ , abstract = {This paper presents a variable bit rate streaming neural speech codec designed for ultra-low bit rate scenarios, based on the SoundStream network framework. The codec employs the vector quantized variational auto-encoder (VQ-VAE) algorithm to capture the temporal structure and spectral characteristics of the speech signal, and constructs a latent space codebook to facilitate the effective mapping of feature vectors to discrete vectors. Based on the harmonic characteristics of speech signals and the inherent defects of single-scale discriminators, we introduce multi-period discriminators and multi-scale discriminators. The training process uses a balanced training strategy to ensure the balance between codebook utilization and training weights, and utilizes the Short-Time Fourier Transform (STFT) spectrum that can provide more accurate time–frequency resolution to compute the reconstruction loss. We introduce codebook loss to improve the utilization rate of the codebook and accelerate the convergence of the model. In the inference process, we use a quantizer selection strategy to achieve adaptive adjustment of variable bitrate. Objective and subjective experiments demonstrate that our proposed new neural speech codec outperforms traditional classical speech codecs and existing neural speech codecs in terms of reconstructed speech naturalness and quality while maintaining the low latency characteristic of neural speech codecs. With a multi-stimulus test with hidden reference and anchor (MUSHRA) score of 87, it is highly suitable for ultra-low bit rate speech compression applications such as satellite speech communication and narrowband instant messaging. The demo has been publicly released at https://svcodec.github.io/ .}, archive = {J_EAAI}, author = {Huaifeng Zhang and Pengfei Wu and Guigeng Li and Yuan An and Hao Zhang}, doi = {10.1016/j.engappai.2025.112418}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112418}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A streaming variable neural speech codec}, volume = {162}, year = {2025}, }
• (2025). LGP-DETR: An end-to-end object detection model for surface defect detection in light guide plates. EAAI, 162, 112416. (www)

  @article{ , abstract = {In this paper, we propose Light Guide Plate-DETR (LGP-DETR), an end-to-end object detection model tailored for identifying surface defects in light guide plates (LGPs). To address challenges such as low contrast, small target size, and complex backgrounds in industrial settings, LGP-DETR integrates three key components: Deformable Transformer Fusion Layer (DtransFusion), a deformable attention-based fusion module for capturing multi-scale features; Upsampling by Dynamic Sampling (DySample), a dynamic upsampling strategy for edge detail preservation; and OrthoC3, a channel attention module that suppresses background noise through orthogonal feature enhancement. We adopt FasterNet as a lightweight convolutional backbone to achieve a balance between accuracy and efficiency. Experimental results on a real-world LGP defect dataset demonstrate that LGP-DETR achieves a mean Average Precision (mAP) of 97.9 % and inference speed of 60 frames per second (FPS), significantly outperforming existing models. Furthermore, generalization tests on a fiberglass fabric defect dataset confirm the model's adaptability to different industrial domains. These findings validate the practical applicability and robustness of the proposed deep learning framework for industrial visual inspection.}, archive = {J_EAAI}, author = {Shuangning Liu and Cunling Liu and Junfeng Li}, doi = {10.1016/j.engappai.2025.112416}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112416}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {LGP-DETR: An end-to-end object detection model for surface defect detection in light guide plates}, volume = {162}, year = {2025}, }
• (2025). Towards efficient stream monitoring: A systematic approach for model selection and continuous improvement in tiny machine learning applications. EAAI, 162, 112415. (www)

  @article{ , abstract = {Measuring ephemeral stream flows is essential for ecological and hydrological studies. However, their intermittent nature and remote locations pose challenges for conventional monitoring methods, which often consume excessive energy to capture rare events. We address this with BODOQUE (Bimodal Observational Device for Optimizing Quantification of Ephemeral streams), a dual-mode system that leverages Tiny Machine Learning (TinyML) on low-power microcontrollers. The system remains in an energy-saving sensing state and activates high-precision measurements only when water flow is detected. We present a model selection methodology that balances detection accuracy with inference cost, enabling reliable operation within hardware constraints. To enhance adaptability in diverse environments, we developed a specialized component that facilitates dataset expansion through new field samples. This supports ongoing retraining to maintain model performance under changing conditions. A comprehensive evaluation using real-world data demonstrates that our system can achieve up to 97% annual energy savings compared to traditional continuous monitoring approaches.}, archive = {J_EAAI}, author = {Benjamín Arratia and Erika Rosas and Javier Prades and Salvador Peña-Haro and José M. Cecilia and Pietro Manzoni}, doi = {10.1016/j.engappai.2025.112415}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112415}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Towards efficient stream monitoring: A systematic approach for model selection and continuous improvement in tiny machine learning applications}, volume = {162}, year = {2025}, }
• (2025). Transformer architecture with illumination aware mechanisms for low-light image enhancement via retinex decomposition. EAAI, 162, 112414. (www)

  @article{ , abstract = {Enhancing low-light images is a complex task that involves not only restoring brightness but also preserving color fidelity and reducing noise interference. In this paper, we propose a novel Retinex-based Transformer Model with Illumination Aware Mechanisms (TIMRetinex-Net), which achieves physically interpretable modeling through a decomposition network guided by Retinex theory. To adapt to light variations in different regions, we randomly apply gamma transformations to several subregions of the illumination component and use a Color Estimation Module to capture the color global distribution of the natural scene in the reflection component. By modeling the color global distribution and repairing the degraded regions collaboratively, we alleviate the issue of being highly sensitive to data usage during training and improve the model’s ability to handle unknown scenes. The Illumination and Reflection Adjustment Transformer Network (IRAT-Net) produces enhanced images, achieving a balanced enhancement of detail and color. In addition, IRAT-Net incorporates an attention mechanism into the feature extraction layer and introduces the Illumination-Guided Information Aggregation Module to adaptively estimate lighting conditions. In the field of image processing, our method based on artificial intelligence was evaluated on five datasets and compared with twelve state-of-the-art methods. The results demonstrated strong alignment with the ground truth, with our method achieving superior performance in both subjective and objective assessments.}, archive = {J_EAAI}, author = {Zixuan Wang and Gang Liu and Hanlin Xu and Yao Qian and Rui Chang and Durga Prasad Bavirisetti}, doi = {10.1016/j.engappai.2025.112414}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112414}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Transformer architecture with illumination aware mechanisms for low-light image enhancement via retinex decomposition}, volume = {162}, year = {2025}, }
• (2025). Efficient interactive segmentation of three-dimensional gaussians with optimal view selection. EAAI, 162, 112413. (www)

  @article{ , abstract = {Three-dimensional (3D) scene representation has advanced rapidly in recent years, drawing the focus of more researchers. One of the main challenges for researchers is quickly and accurately segmenting 3D objects. Previous work has achieved excellent segmentation accuracy, but retraining requires a significant amount of time. Additionally, most methods fail to provide users with an efficient and convenient segmentation experience. To address these issues, we present Efficient Interactive Segmentation of 3D Gaussians (EISG), an efficient interactive segmentation method that eliminates the need for lengthy retraining. We first design an optimal view selection (OVS) method. This method uses 3D Gaussian entropy and image uncertainty to evaluate the quantity of view information. OVS helps users quickly select the optimal segmentation view, thereby enhancing interaction efficiency. Secondly, we use projection to find the target foreground rapidly and then segment the approximate objects using a clustering algorithm. Thirdly, we design a spatial-color background filter (SCBF) using the depth and color of 3D Gaussians. SCBF enables precise segmentation without needing retraining. Our method has been systematically tested on multiple datasets. Compared to other methods, the results demonstrate that EISG achieves ideal accuracy while significantly reducing processing time.}, archive = {J_EAAI}, author = {Yongtang Bao and Chengjie Tang and Yuze Wang and Yutong Qi and Ruijun Liu}, doi = {10.1016/j.engappai.2025.112413}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112413}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Efficient interactive segmentation of three-dimensional gaussians with optimal view selection}, volume = {162}, year = {2025}, }
• (2025). Adversarial black-box attack and defense for convolutional neural network-based power quality disturbance classification. EAAI, 162, 112411. (www)

  @article{ , abstract = {Correctly identifying power quality disturbance (PQD) is crucial for the proper functioning of power systems. Deep learning (DL) techniques have been widely used for PQD classification due to their excellent performance. However, DL models are susceptible to adversarial attacks, posing a serious security threat to DL-based PQD classification systems. This issue has received limited attention in current research. In this study, we first utilize a convolutional neural network (CNN) to recognize various types of PQD signals. To evaluate model robustness, we introduce a black-box attack method for PQD classification based on the variance-tuning momentum iterative fast gradient sign method (VMI-FGSM). VMI-FGSM integrates a variance tuning method into the iterative process of the momentum iterative fast gradient sign method (MI-FGSM) , thereby producing more transferable adversarial PQD signals. To defend against such attacks, we propose a perturbation removal defense based on a generative adversarial network (PRD-GAN). This approach is capable of removing perturbations from adversarial PQD signals before they are recognized by the target classification model. Experiments demonstrate that VMI-FGSM produces adversarial perturbations that are nearly identical to those of the advanced MI-FGSM, but its adversarial examples are significantly more effective at misleading the target CNN model. Furthermore, the proposed PRD-GAN effectively reconstructs adversarial PQD signals into clean forms under various black-box attack intensities and outperforms the multi-level denoising autoencoder (ML-DAE) in defense performance due to its superior reconstruction capability.}, archive = {J_EAAI}, author = {Xiudong Zhang and Congmei Jiang and Mingbiao Yu and Xiankui Wen and Jing Zhang and Na Rong and Song Han}, doi = {10.1016/j.engappai.2025.112411}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112411}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adversarial black-box attack and defense for convolutional neural network-based power quality disturbance classification}, volume = {162}, year = {2025}, }
• (2025). An intrusion detection system for critical infrastructures: Modbus approach. EAAI, 162, 112410. (www)

  @article{ , abstract = {This study aims to develop an Intrusion Detection System (IDS) using deep learning and machine learning algorithms to detect cyber attacks in the network traffic of critical infrastructures using an artificial intelligence-based approach. The research investigates various machine learning algorithms, datasets, and performance evaluations to detect the security vulnerabilities commonly found in industrial networks. Implemented in Python, the system has been tested on hybrid dataset, demonstrating the performance of different algorithms in terms of accuracy, precision, and other metrics. From artificial intelligence perspective, this study contributes machine learning and deep learning in cybersecurity, showing how normal and ensemble models can effectively detect complex threats, with fewer features but more relevant. The research employs supervised learning techniques, leveraging labeled datasets to train models that can accurately classify network traffic as either normal or attack, ensuring high detection accuracy. From an engineering standpoint, the system’s Python implementation addresses the practical challenges of real-world deployment in industrial control systems (ICS) and facilitates integration with existing infrastructures. Additionally, the custom dataset and post-dissector code contribute to the field of industrial cybersecurity, providing engineers with tools for testing, validating, and optimizing IDS solutions. As cyber–physical systems are increasingly integrated into ICS, the proposed IDS provides a crucial layer of defense against cyber threats, safeguarding both the digital and physical components of critical infrastructure. The findings reveal that the proposed system exhibits high performance in terms of detection accuracy. The results show that the system provides an effective and reliable detection mechanism using artificial intelligence techniques.}, archive = {J_EAAI}, author = {Murat Varol and Murat İskefiyeli}, doi = {10.1016/j.engappai.2025.112410}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112410}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An intrusion detection system for critical infrastructures: Modbus approach}, volume = {162}, year = {2025}, }
• (2025). Adaptive resource management in dynamic Cyber–Physical systems using artificial intelligence. EAAI, 162, 112409. (www)

  @article{ , abstract = {Effective resource management is crucial for the operation of dynamic Cyber–Physical Systems (CPS), yet traditional static or rule-based approaches often fail to handle their inherent complexity. This paper presents a novel Artificial Intelligence (AI)-driven framework for adaptive resource management, defined as the capability to autonomously adjust resource allocation by proactively forecasting future demands and dynamically optimizing decisions in real-time. The framework integrates a suite of AI techniques, including time-series models like Long Short-Term Memory (LSTM), chosen for their ability to capture complex temporal dependencies, for demand prediction. For resource allocation, it employs advanced actor–critic reinforcement learning (RL) algorithms like Proximal Policy Optimization (PPO) and Deep Deterministic Policy Gradient (DDPG), selected for their stability and efficiency in complex decision-making tasks. Performance was rigorously evaluated in a simulated dynamic environment. Experimental results demonstrate that combinations leveraging LSTM’s predictive accuracy with the robust optimization of PPO and DDPG achieve superior performance and stability. Specifically, the LSTM+DDPG and LSTM+PPO configurations yielded the highest average rewards (0.964 and 0.942, respectively), significantly outperforming the fixed-strategy baseline (0.497) and other AI pairings. Furthermore, the feasibility of training prediction models in a distributed manner via Federated Learning (FL) is successfully demonstrated. This research highlights that a synergistic integration of suitable AI predictors and advanced RL agents provides a powerful and resilient solution for resource management in dynamic CPS.}, archive = {J_EAAI}, author = {Xiaofei Zhao and Fangling Guo and Amin Huang and Jieqiong Ding and Chi Yan and Wei Yuan and Yunqi Su and Quanzhou Li and Qianggang Zhang}, doi = {10.1016/j.engappai.2025.112409}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112409}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive resource management in dynamic Cyber–Physical systems using artificial intelligence}, volume = {162}, year = {2025}, }
• (2025). A novel prior knowledge-based and texture-enhanced network for few-shot industrial defect segmentation. EAAI, 162, 112408. (www)

  @article{ , abstract = {Deep learning-based segmentation techniques have demonstrated significant potential in defect detection, which is vital for product quality. However, the existing models tend to specialize in detecting specific defect types, reducing their adaptability to a wider range of product defects, most existing methods rely on multi-scale or prototype learning to extract defect features, but still struggle with complex backgrounds, various interferences, and large intra-class variations. Additionally, the rarity of certain defects limits the availability of training samples. Herein, an innovative segmentation network, the Prior Knowledge-based and Texture-Enhanced Network (PTNet), is designed for few-shot industrial segmentation. The model is mainly composed of a self-guidance branch, a cross-guidance branch, and a texture enhancement module, enabling generalization across defect types with minimal labeled samples. Self-guidance extracts prior knowledge from the query image, while the cross-guidance branch extracts prior and prototype features from the masked support image. The texture information in the low-level features of the backbone is then enhanced by proposed texture enhancement module (TEM). Finally, the enhanced low-level texture information are fused with high-level semantic features, allowing the network to fully exploit both local details and global context before being decoded to restore the original image size. This enables the model to handle complex textures and generalize to unseen defect types. Extensive experimental results validate the effectiveness of the proposed modules. State-of-the-art performance is achieved in few-shot defect segmentation, with notable improvements in mean Intersection over Union (mIoU) of 46.05 % and 46.98 % under 1-shot and 5-shot conditions, respectively.}, archive = {J_EAAI}, author = {Xingyue Liu and Qian Wu and Yahui Cheng and Guojun Wen}, doi = {10.1016/j.engappai.2025.112408}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112408}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel prior knowledge-based and texture-enhanced network for few-shot industrial defect segmentation}, volume = {162}, year = {2025}, }
• (2025). A novel reinforcement learning-based approach for optimal control: An application to multi-tank water systems. EAAI, 162, 112407. (www)

  @article{ , abstract = {The optimization of control actions is a critical challenge in industrial systems, especially when dealing with complex and unknown dynamics. Data collected from the environment enables the application of reinforcement learning techniques, which let the controller learn a policy based on data. This work proposes a novel model-free reinforcement learning approach that consists of a value iteration algorithm based on separate policy evaluation and policy improvement phases to provide an accurate control policy estimation. The proposed approach addresses tracking control for quadruple-tank water systems while obtaining minor tracking errors and faster transient responses. The results from the case study reveal better accurate estimation of the value function, up to 86.13% mean improvement in tracking accuracy and faster responses compared to existing methods. Therefore, the proposed approach demonstrates advantages in optimizing control performance and stands as a promising control method for industrial applications.}, archive = {J_EAAI}, author = {Eva Masero and Giacomo Mussita and Alessio La Bella and Riccardo Scattolini}, doi = {10.1016/j.engappai.2025.112407}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112407}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel reinforcement learning-based approach for optimal control: An application to multi-tank water systems}, volume = {162}, year = {2025}, }
• (2025). Automatic weighted ensemble learning for ballast resistance estimation driven by vehicle-ground information fusion. EAAI, 162, 112406. (www)

  @article{ , abstract = {Health management of transmission parameters for railway signal equipment is a key link between intelligent operation and maintenance. As a core parameter of track circuits, ballast resistance significantly affects signal transmission. To accurately and reliably assess its health state, an ensemble learning algorithm (ELA) is introduced, tackling deviations of appraisal decision boundaries. Focusing on issues of complex weight calculation, model homogenization, and severe overfitting in ELA, an integration model based on an automatic weight allocation strategy (AWAS) is innovatively proposed, constructing a method for resistance estimations driven by information fusion, while maximizing its generalization ability. Firstly, for deterioration mechanism analysis of ballast resistance, a transmission state model for vehicle-ground collaboration is established, completing extractions of evolutionary rules. Secondly, the improved ELA leverages heterogeneous classifier optimization and automatic weighted soft voting, with its core ensemble strategy employing a secondary learner to map the fused datasets. Then, by means of data mining techniques, interpolation and denoising algorithms are applied to implement data preprocessing, facilitating the effective fusion of heterogeneous vehicle-ground information. Finally, based on occurrence of adverse conditions, an appropriate particle size is set to achieve state warning. The results indicate that the proposed AWAS for ballast resistance calculations can achieve 98.52 % testing accuracy and outperforms others.}, archive = {J_EAAI}, author = {Conghui Wang and Shiwu Yang and Chang Liu}, doi = {10.1016/j.engappai.2025.112406}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112406}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Automatic weighted ensemble learning for ballast resistance estimation driven by vehicle-ground information fusion}, volume = {162}, year = {2025}, }
• (2025). Future of humanity in an artificial intelligence centric world. EAAI, 162, 112405. (www)

  @article{ , abstract = {This scholarly article rigorously investigates the transformative and disruptive roles of artificial intelligence (AI) in influencing the trajectory of human society. By concentrating on three fundamental sectors—healthcare, finance, and education—it evaluates the ways in which AI augments operational efficiency, facilitates intricate decision-making processes, and introduces innovative capabilities such as personalized medicine and automated financial systems. Concurrently, the analysis underscores urgent ethical dilemmas, encompassing algorithmic bias, accountability deficiencies, data privacy vulnerabilities, and workforce displacement. Employing real-world examples such as Deepfakes, and Neuralink, the article contextualizes emerging challenges within a dynamic socio-technical framework. The research offers a cohesive conceptual model that amalgamates technical, ethical, and governance aspects of AI, while presenting policy recommendations designed to promote transparency, equity, and human-centered AI development. The study emphasizes the necessity for reliable AI systems that humans can trust. The conclusions accentuate the immediate necessity for robust regulatory frameworks and sector-specific ethical supervision to ensure that advancements in AI are harmonized with the well-being of society.}, archive = {J_EAAI}, author = {Sunil Baloda and Monika Sharma and Mukesh Kumar}, doi = {10.1016/j.engappai.2025.112405}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112405}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Future of humanity in an artificial intelligence centric world}, volume = {162}, year = {2025}, }
• (2025). A novel federated deep learning for intrusion detection in smart grid cyber-physical systems. EAAI, 162, 112404. (www)

  @article{ , abstract = {The fusion of sophisticated computational, communicative, and physical elements in Smart Grid Cyber-Physical Systems (SGCPS) has greatly improved the efficiency and reliability of power grids. However, this complexity introduces enhanced cybersecurity risks, evidenced by significant cyberattacks on the Ukrainian power grid during 2015 and 2016. Despite progress in Artificial Intelligence (AI)-driven security solutions for SGCPS, practical deployment of these technologies is often limited due to a lack of high-quality attack data and owners’ hesitance to distribute sensitive details. This paper introduces an innovative strategy to fortify SGCPS against diverse network threats via a comprehensive intrusion detection system. We present a deep learning model leveraging a temporal convolutional network with multi-feature integration, aimed at robust threat identification. We also propose a federated learning framework enabling various SGCPS to jointly develop an extensive intrusion detection model, ensuring data privacy. Moreover, we incorporate a gradient compression technique utilizing the Long Short Term Memory- β -Total Correlation Variational Autoencoder (LSTM- β -TCVAE) model to enhance and secure model parameters throughout the training phase. Thorough experimental validations confirm the efficacy of our method in recognizing multiple cyber threat types to SGCPS and its advantages over current methods.}, archive = {J_EAAI}, author = {Rong Xie and Bin Wang and Xin Xu}, doi = {10.1016/j.engappai.2025.112404}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112404}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel federated deep learning for intrusion detection in smart grid cyber-physical systems}, volume = {162}, year = {2025}, }
• (2025). Tunnel-YOLO: An improved you only look once algorithm for real-time shield tunnel lining leakage detection. EAAI, 162, 112403. (www)

  @article{ , abstract = {Detecting crack leakages in shield tunnels is crucial for ensuring structural safety and extending service life, as traditional detection methods are limited by high subjectivity and low accuracy. To address these limitations, this paper proposes Tunnel-YOLO, an improved object detection algorithm based on You Only Look Once version 8 (YOLOv8). This algorithm replaces standard convolutional blocks with a novel Receptive Field Channel Attention Convolution (RFCAConv) module, which leverages dynamic receptive fields to enhance feature capture at different scales. We also introduce a C2f_SGE module, integrating the Spatial Group-wise Enhance (SGE) attention mechanism into the C2f (CSPNet with 2 convolutions) block to significantly improve feature extraction while suppressing background interference. Furthermore, an Edge Feature Enhancement Detection Head (EFE-Head) incorporates deconvolution layers to enhance fine-grained details for more precise boundary localization. To better accommodate the shape-sensitive detection task, our LeShape-IoU (Intersection over Union) loss function is designed to focus on the shape and scale characteristics of target bounding boxes. Experimental results on a public, real-world dataset demonstrate that Tunnel-YOLO significantly outperforms the baseline, increasing Recall, Precision, and mean Average Precision at 0.5 IoU (mAP50) by 15.7%, 10.3%, and 14.8%, respectively. Comparative analysis with other mainstream algorithms further validates the effectiveness and superiority of the proposed Tunnel-YOLO.}, archive = {J_EAAI}, author = {Ruijun Yang and Hao Zhang}, doi = {10.1016/j.engappai.2025.112403}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112403}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Tunnel-YOLO: An improved you only look once algorithm for real-time shield tunnel lining leakage detection}, volume = {162}, year = {2025}, }
• (2025). Development of data-driven predictive model and enhanced multiobjective optimization to improve the excavation performance of large-diameter slurry shields. EAAI, 162, 112402. (www)

  @article{ , abstract = {Safety, efficiency and energy consumption are important aspects for evaluating the performance of large-diameter slurry shield, and improving the performance of shield is crucial for safe and efficient excavation. To this end, a data-driven hybrid method is developed to improve the excavation performance of large-diameter slurry shields by intelligence regulating shield parameters. This method combines Bayesian Optimization with categorical boosting (BO-CatBoost) and enhanced multiobjective evolutionary algorithm based on decomposition (EMOEA/D). The method uses surface settlement, penetration and specific energy as output targets and employs the expert knowledge to select the input parameters. Subsequently, the trained BO-CatBoost model is employed to fit the input-output relationship. On this basis, the multiobjective optimization process was performed using EMOEA/D, with the important parameters determined by Shapley Additive exPlanations as decision variables and the nonlinear relationship fitted by BO-CatBoost as the objective function. Finally, the technique for order preference similarity to ideal solution is applied to obtain optimal operational parameters, thereby enhancing the excavation performance of large-diameter slurry shield. The proposed method is applied to a Wuhan rail transit line to verify the effectiveness, and the result shows that: (1) Our method can accurately predict the three targets with goodness of fit ranging from 0.938 to 0.988, respectively. (2) The proposed method can effectively improve the excavation performance of the large-diameter slurry shield, and reaches 13.88 %, 5.21 %, and 10.88 %, respectively. (3) An adaptive decision-making system for setting operational parameters is constructed, which is valuable for formulating of operational control strategies for large-diameter slurry shields.}, archive = {J_EAAI}, author = {Feiming Su and Xianguo Wu and Tiejun Li and Yang Liu}, doi = {10.1016/j.engappai.2025.112402}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112402}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Development of data-driven predictive model and enhanced multiobjective optimization to improve the excavation performance of large-diameter slurry shields}, volume = {162}, year = {2025}, }
• (2025). Contrastive active adaptive partial label learning under class distribution mismatch. EAAI, 162, 112401. (www)

  @article{ , abstract = {Partial label learning is an important learning framework where each training sample is associated with a candidate label set and its ground-truth label is included in the candidate label set. Active partial label learning is a variation where training data consists of both labeled and unlabeled samples. However, there exists the problem of class distribution mismatch, wherein the unlabeled sample set contains many instances out of the target categories. In this paper, a contrastive active adaptive partial label learning method under class distribution mismatch which combines active partial label learning with contrastive coding is proposed. A novel active sample selection strategy is first established to use label propagation ability to measure the optimization ability of unlabeled samples to partially labeled samples. Furthermore, to solve the problem of class distribution mismatch, a joint query score based on contrastive coding is utilized to reduce the queries of unlabeled samples out of target categories. Finally, the above two indicators are combined adaptively to select the most valuable unlabeled samples in target categories for manual labeling and the selected samples will be added to the training sample set to train the new classifier. The effectiveness and efficiency of the method are evaluated by performing experiments on the datasets CIFAR10 and CIFAR100.}, archive = {J_EAAI}, author = {Aohan Zhang and Kezhen Dong and Hongying Zhang}, doi = {10.1016/j.engappai.2025.112401}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112401}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Contrastive active adaptive partial label learning under class distribution mismatch}, volume = {162}, year = {2025}, }
• (2025). A framework for photovoltaic power forecasting based on hybrid data reconstruction, neural network models fusion, and multi-objective optimization. EAAI, 162, 112400. (www)

  @article{ , abstract = {The intermittency of photovoltaic power seriously affects the safety and operation of the power grid. Accurate photovoltaic power forecasting is critical for a safe connection of large-scale solar energy to the grid. Despite the efforts of many researchers, current forecasting methodologies remain inadequate. To bridge this gap, leveraging the recent advancements in artificial intelligence algorithms, this work combines cutting-edge deep learning techniques and data preprocessing strategies to develop a forecasting system that comprehensively considers various influencing factors and integrates multiple deep learning neural networks. The framework enables deterministic forecasting and uncertainty analysis, providing reliable supporting information for accurate forecasting through hybrid decomposition data preprocessing and feature selection modules. Then, closed-form continuous-time (Cfc) neural networks are introduced as one of the core forecasting components. Theoretically, the validity of the combined model and the Pareto optimization process are proved. Practically, the multi-objective African vultures optimization (MoAvo) is employed to identify the Pareto optimal solution, integrate four models, and improve the model's adaptability to external environmental changes. The experimental results show that the average mean absolute percentage error (MAPE) of the designed combined system for 1–3 steps forecasting on the Yulara are 6.09 %, 8.15 %, and 10.03 %, respectively. The results demonstrate that the framework fully considers the influence of candidate variables on forecasting, offering significant advantages over comparison models.}, archive = {J_EAAI}, author = {Menggang Kou and Jianzhou Wang and Jingrui Li and Runze Li and Zhiwu Li}, doi = {10.1016/j.engappai.2025.112400}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112400}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A framework for photovoltaic power forecasting based on hybrid data reconstruction, neural network models fusion, and multi-objective optimization}, volume = {162}, year = {2025}, }
• (2025). A binary particle swarm optimization with dual encoding mechanism for feature selection. EAAI, 162, 112397. (www)

  @article{ , abstract = {Feature selection is a crucial machine learning preprocessing stage with numerous practical uses. Numerous algorithms are created to tackle the task. However, these algorithms still suffer from several challenges. This study introduces a novel binary particle swarm optimization with dual encoding mechanism, named DEBPSO, to solve feature selection problems. A new transfer function, called an inverse S-shaped function, and a Boolean encoding mechanism are employed to enhance the exploration performance of DEBPSO. In addition, to better balance exploration and exploitation, a new game mechanism is proposed to find the best solution. In order to verify the performance of DEBPSO, a comprehensive experimental is designed. The experimental results on 27 well-known datasets show that DEBPSO significantly outperforms the compared algorithms on 17, 14, 13, 17, 16, 21, 14, 23, 20 datasets in terms of classification error rate, highlighting its efficiency in reducing the classification error rate and irrelevant features.}, archive = {J_EAAI}, author = {Chong Zhou and Rumeng Liang and Qi Liu and Sirui Niu}, doi = {10.1016/j.engappai.2025.112397}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112397}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A binary particle swarm optimization with dual encoding mechanism for feature selection}, volume = {162}, year = {2025}, }
• (2025). A novel weight-optimized machine-learning hybrid model for daily river runoff prediction. EAAI, 162, 112396. (www)

  @article{ , abstract = {The daily runoff process has been characterized as nonlinear and unsteady due to the impacts of watershed precipitation and evaporation, vegetation coverage rate, reservoir operations and other human activities. In recent years, machine-learning (ML) models have been widely applied in the daily runoff predictions, but the robustness and effectiveness of individual ML model is always limited. A novel weight optimization scheme has been introduced to ML models to obtain accurate predictions of daily river runoff. Variational modal decomposition method is adopted in the dataset preprocessing, and the runoff prediction performance of various classic ML models, including Genetic Algorithm-Back Propagation neural network (GA-BP), Long Short-Term Memory network (LSTM), Elman neural network (Elman) and Genetic Algorithm-Support Vector Machine (GA-SVM) are subsequently evaluated. A particle swarm optimization (PSO) based weight optimization strategy is proposed to combine different types of ML models, thus more accurate and robust results could be obtained. The ten-fold cross-validation method has been adopted and the performance of the optimized hybrid models are further evaluated for different schemes. A case study at Hankou hydrological station demonstrates that root mean square error (RMSE) and mean absolute percentage error (MAPE) is improved by 35.7 %, 75.8 % respectively for the optimized hybrid model. The present study shares useful insights to the comprehensive optimization of various ML models in the intelligent management of water resources.}, archive = {J_EAAI}, author = {Zhonglian Jiang and Jianglong Ying and Zhen Yu and Xiao Chu and Chengqiang Yu}, doi = {10.1016/j.engappai.2025.112396}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112396}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel weight-optimized machine-learning hybrid model for daily river runoff prediction}, volume = {162}, year = {2025}, }
• (2025). Intelligent assessment of habitat quality based on multiple machine learning fusion methods. EAAI, 162, 112395. (www)

  @article{ , abstract = {Evaluating habitat quality can help balance the relationship between economic development and biodiversity conservation, and it serves as a foundation for constructing an ecological security pattern. However, research on the intelligent construction of habitat quality is limited. This study develops a comprehensive framework to assess habitat quality based on optimized machine learning methods. The findings of the research are as follows: (1) From the perspective of human-machine interactive interpretation, ensemble learning is used to enhance the performance of basic classifiers, resulting in a classification map with high precision and recall. (2) The particle swarm optimization (PSO) algorithm can improve the goodness of fit of the Extreme Gradient Boosting (XGBoost) inversion model by 4–5 %. (3) The habitat quality inversion method based on XGBoost-PSO has high credibility and application value, with its texture structure being the result of both expert experience and image information interaction. (4) The model demonstrates certain application potential in downscaling; under the seven-band perspective, the blue and near-infrared bands are the most important, while in the four-band perspective, green and near-infrared bands take precedence.}, archive = {J_EAAI}, author = {Kui Yang and Dongge Cui and Chengrui Wang and Qi Tang and Linguang Miao}, doi = {10.1016/j.engappai.2025.112395}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112395}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Intelligent assessment of habitat quality based on multiple machine learning fusion methods}, volume = {162}, year = {2025}, }
• (2025). Hierarchical semantics guided multi-scale correlation network for alignment-free red-green-blue and thermal salient object detection. EAAI, 162, 112394. (www)

  @article{ , abstract = {RGBT (red-green-blue and thermal) salient object detection (SOD) aims to identify and highlight the most visually salient objects in an image by leveraging the complementary information from both RGB and thermal (TIR) modalities. It is particularly effective for 24/7 intelligent surveillance and autonomous perception in smart city security and traffic monitoring, especially under low light and adverse weather. However, existing methods primarily rely on manually aligned datasets, which are limited in handling the challenges posed by unaligned multi-modal data in real-world applications. Furthermore, these methods usually extract complementary information from both modalities using fixed-size windows (Liuet al., 2022, Wanget al., 2024b). However, such fixed-size windows are not effective in dealing with unaligned multi-modal images due to spatial inconsistencies. Additionally, existing methods often use single-layer high-level feature to represent semantic information, which fails to fully exploit the complementary benefits of multi-level features, thereby reducing the effectiveness of semantic guidance. To address these challenges, we propose a Hierarchical Semantics guided Multi-scale correlation Network (HSMNet) for alignment-free RGBT SOD. A Hierarchical Semantic Fusion Module (HSFM) dynamically assigns weights to features from multiple levels, enabling adaptive fusion of multi-level semantic information. A Multi-scale Asymmetric Correlation Module (MACM) employs windows of various sizes to capture asymmetric correlations between unaligned multi-modal data, enhancing cross-modal complementary information extraction even when data are not perfectly aligned. We conduct extensive experiments on unaligned, weakly aligned and aligned RGBT SOD datasets, with results demonstrating that our method outperforms state-of-the-art algorithms, achieving superior accuracy and robustness in both unaligned and weakly aligned RGBT SOD scenarios.}, archive = {J_EAAI}, author = {Chengmei Han and Lei Liu and Kunpeng Wang and Fei Xie and Bing Wei}, doi = {10.1016/j.engappai.2025.112394}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112394}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hierarchical semantics guided multi-scale correlation network for alignment-free red-green-blue and thermal salient object detection}, volume = {162}, year = {2025}, }
• (2025). Nonlinear dynamic modeling of turbojet engines using combined convolutional and long short-term memory networks. EAAI, 162, 112393. (www)

  @article{ , abstract = {Turbojet engines are widely used in small-scale aerial vehicles, but their nonlinear and time-varying dynamics present significant challenges for accurate modeling and control. Traditional system identification methods often struggle to capture these complex behaviors, particularly under limited data conditions. This study proposes a novel hybrid neural network architecture that combines convolutional neural networks and long short-term memory units. The model is specifically designed for small-sample scenarios, enabling robust learning and precise engine speed prediction from real input-output sequences. The input vector comprises the current engine speed, the next-step pulse-width modulation command, and its increment, enhancing the model’s responsiveness and reducing phase-lag effects. The proposed model is trained and evaluated on a real-world dataset containing 38,257 samples, with 80 % used for training and 20 % for testing. Its predictive performance is assessed using step input responses and three evaluation metrics: mean absolute error, root mean square error, and Pearson correlation coefficient. Experimental results demonstrate that the proposed hybrid architecture outperforms other recurrent models in capturing transient dynamics and accurately reproducing real engine behavior. These findings highlight the model’s effectiveness in modeling nonlinear engine dynamics and its potential as a data-efficient alternative to traditional identification techniques for small-scale turbojet applications.}, archive = {J_EAAI}, author = {Chen Lei and Dong Wei and Su Hang and Chi Yutian and Tian Congling and Gao Yongzhuo and Wu Dongmei and Dong Hui}, doi = {10.1016/j.engappai.2025.112393}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112393}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Nonlinear dynamic modeling of turbojet engines using combined convolutional and long short-term memory networks}, volume = {162}, year = {2025}, }
• (2025). Scientific machine learning for generic compact model parameter extraction of nanoscale transistors. EAAI, 162, 112392. (www)

  @article{ , abstract = {This work presents a framework to automate modelcard extraction of industry-standard compact models. This framework presents a Scientific Machine Learning (ScML) approach capable of inverse modeling. It integrates a random forest model with an artificial neural network to produce efficient and precise regression results. This new method is used for parameter extraction in the standard compact models of the semiconductor device industry. Modeling of a multi-gate Field Effective Transistor (FET) with an industry-standard compact model like Berkeley Short-channel Insulated-Gate Field-Effect Transistor Model – Common Multi-Gate (BSIM-CMG) is taken as an example to illustrate and describe the framework and highlight its key advantages. Proposed framework is useful in numerous aspects; it holds vital principles of physics, avoids depending on massive datasets, and has a sparse architecture while avoiding accuracy trade-offs. The framework is tested on production-level experimental devices to evaluate the real-world performance. This framework significantly reduces the time and cost of parameter extraction for the Process Design Kits (PDKs) development. Therefore, this is of immediate importance for fabrication and Electronic Design Automation (EDA) industries.}, archive = {J_EAAI}, author = {Kumar Sheelvardhan and Surila Guglani and Abhilash Dubey and Shashank Dubey and Sindhu Ramaswamy and Vaidy Subramanian and Kassandra Anderson and Glenn Workman and Sourajeet Roy and Avirup Dasgupta}, doi = {10.1016/j.engappai.2025.112392}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112392}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Scientific machine learning for generic compact model parameter extraction of nanoscale transistors}, volume = {162}, year = {2025}, }
• (2025). Physics-informed graph neural network for 3D spatiotemporal structural response modeling of flexible pavements. EAAI, 162, 112391. (www)

  @article{ , abstract = {Quantifying pavement damage is crucial for roadway agencies' maintenance planning. This study proposed a Physics-informed Graph Neural Network-based Pavement Simulator (PhyGPS) to predict three-dimensional (3D) asphalt concrete pavement responses, building upon an established data-driven Graph Neural Network-based Pavement Simulator (GPS) model. The key innovation lies in integrating knowledge graphs and mechanics equations to create a physics loss function, distinguishing it from its data-driven counterpart. The physics loss function comprises strain-displacement and stress loss components derived from 3D strain-displacement relations and stress equilibrium principles. A thorough 3D finite element (FE) pavement database supported the model development. The 3D FE pavement data was transformed into graph format where nodes and edges represent 3D FE pavement models’ nodes and node connections, respectively. Performance evaluation employed two case studies: “OneStep” for assessing short-term predictive capabilities and “Rollout” for examining long-term prediction accuracy under practical conditions. Results demonstrated that the physics-informed GPS model showed superior long-term predictive capability and robustness while maintaining excellent short-term accuracy compared to the data-driven model. Both models achieve rollout time under 8 s per FE simulation case, a dramatic improvement over the 12-h runtime of traditional 3D FE pavement models. The PhyGPS model successfully integrates physics principles, spatial relationships between structural components, temporal correlations in structural data, and complex material properties, offering an accurate, robust, and computationally efficient solution for predicting 3D pavement responses.}, archive = {J_EAAI}, author = {Fangyu Liu and Imad L. Al-Qadi}, doi = {10.1016/j.engappai.2025.112391}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112391}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Physics-informed graph neural network for 3D spatiotemporal structural response modeling of flexible pavements}, volume = {162}, year = {2025}, }
• (2025). Health state estimation of retired batteries based on physical constraints. EAAI, 162, 112390. (www)

  @article{ , abstract = {With the growing use of retired lithium-ion batteries, accurately monitoring their health status has become increasingly important. This study proposed a method to estimate the health of retired batteries by embedding their capacity degradation characteristics directly into the loss function of a Bidirectional Long Short-Term Memory (BiLSTM) network, combined with a Physically Informed Neural Network (PINN) model. The model is developed by incorporating the dynamics of the solid electrolyte interface (SEI) membrane, which evolves as the lithium-ion poles of the retired battery move. By combining these dynamics with the governing equations of motion, a partial differential equation (PDE) is derived. This approach integrates physical constraints, data-driven learning, and PDEs into a composite loss function. The proposed method is validated on two different datasets under varying operating temperatures. The results show that the PINN-BiLSTM model achieves a Root Mean Square Percentage Error (RMSPE) of 0.024, representing a 9.67 % improvement over the PINN-LSTM. This adaptive PINN method offers highly accurate health state predictions across temperature variations, thus supporting the sustainable use of retired batteries in secondary applications and helping to mitigate energy scarcity.}, archive = {J_EAAI}, author = {Fei Xia and Qianwen Dong and Lin Xia and Zhenyi An and Ziyang Xia and Chunyang Gong}, doi = {10.1016/j.engappai.2025.112390}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112390}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Health state estimation of retired batteries based on physical constraints}, volume = {162}, year = {2025}, }
• (2025). Adaptive cybersecurity in bio-medical modules: A dynamic programming neural network-based protection for markov-chain fabricated data attacks. EAAI, 162, 112389. (www)

  @article{ , abstract = {With the advances in communication technologies, remote monitoring and control of drug delivery are becoming prevalent in bio-medical engineering. In a chemotherapy system, the measurement signals can be wirelessly transferred to the control center by communication networks. Nevertheless, the avenues of communication might be jeopardized by false data injection (FDI), which poses significant risks to the security and stability of biomedical systems. In this work, a defense mechanism is developed to tackle the effect of FDI threats in the networked chemotherapy system. In particular, the effect of FDI attacks on the chemotherapy system is modeled by the Markov chain process. The proposed defense mechanism is designed in two parts: i ) a data-driven sliding mode observer (DDSMO) is utilized to identify the occurrence of cyber attacks in the tumor signals measured by the bio-sensor, and ii ) a mitigation scheme based on a dynamic rejection compensator (DRC) to compensate for the impact of cyber threats. In the mitigation phase, a goal representation heuristic dynamic programming (GrHDP) is adopted to adaptively adjust the parameters of DRC and to dynamically handle the cyber threats. The designed mitigation mechanism not only regulates the cancerous cells against cyber threats but also minimizes the side effects of drug delivery by regulating the output of normal cell and immune cell. Compared to prevalent methodologies, the proposed approach yields significant performance, including a 60.23 % improvement over the without protection, 37.48 % over the DDSMO-based model predictive controller (MPC), 35.95 % over the reinforcement learning (RL) based Kalman filter, and 70.44 % over the proportional integral (PI) based Kalman filter.}, archive = {J_EAAI}, author = {Mostafa Taheri and Juliang Yin and Zahra Rasooli Berardehi}, doi = {10.1016/j.engappai.2025.112389}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112389}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive cybersecurity in bio-medical modules: A dynamic programming neural network-based protection for markov-chain fabricated data attacks}, volume = {162}, year = {2025}, }
• (2025). Weighted multi-source domain unsupervised adaptive network for rotating machinery fault diagnosis based on dual adversarial. EAAI, 162, 112386. (www)

  @article{ , abstract = {Multi-domain adaptive methods are becoming a growing focus of fault diagnosis, which can provide enhanced data support for models using the feature information from various source domains. As a most commonly used method, unsupervised multi-domain adaptive methods (UMA) can eliminate the requirement for the label of the target domain samples. However, the neglect of contributions from different source domains to the target domain and insufficient utilization of diagnostic information from multiple source domains are the widely limitation of UMA. Therefore, a dual-adversarial weighted multi-source domain unsupervised adaptive network (DAWMUN) is proposed to utilize diagnostic information from multi-source domains and consider the contribution of different source domains. Firstly, the shared feature extractor and dual adversarial training with the domain adversarial modules between multi-source domains and source-target domains are used to enhance domain confusion between multi-source and target domains (MSTD). Secondly, based on Multiple Kernel Maximum Mean Discrepancy (MK-MMD), a novel weighting mechanism and the corresponding training framework are constructed to effectively reduce negative transfer. Finally, a novel weighted classifier is proposed to merge the outputs of multiple classifiers and synthesize the impact of each source domain. The performance of the DAWMUN is validated using a rotating machinery dataset across various transfer tasks under different rotational speed and load conditions. The experimental results demonstrate that the diagnostic accuracy using the proposed DAWMUN is superior to existing SSDA and MSDA methods, with the average accuracies of 98.53 % and 98.23 % across six tasks in two separate experimental setups. The comparison to the existing methods results that the DAWMUN still demonstrates superior performance with improvements of 2.54 % and 2.86 %, respectively.}, archive = {J_EAAI}, author = {Wenqi Wang and Zongzhen Zhang and Jinrui Wang and Baokun Han and Huaiqian Bao and Zhikang Fan and Rongkang Ge}, doi = {10.1016/j.engappai.2025.112386}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112386}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Weighted multi-source domain unsupervised adaptive network for rotating machinery fault diagnosis based on dual adversarial}, volume = {162}, year = {2025}, }
• (2025). Twin-stream network: Enhancing wave prediction by capturing spatiotemporal information. EAAI, 162, 112385. (www)

  @article{ , abstract = {Wave prediction is a critical challenge in ocean and coastal engineering, particularly for understanding and mitigating the effects of sea waves on structures such as ships, offshore platforms, and coastal defenses. A novel machine learning model, Twin-Stream Network (TSNet), is proposed to enhance wave prediction accuracy by leveraging temporal and spatial dependencies in historical data. The TSNet model along with other baseline models are evaluated, in both single-point and multi-point forecasting tasks, by various performance metrics across different datasets including one-dimensional-linear, one-dimensional-nonlinear, two-dimensional-linear and two-dimensional-nonlinear water waves. The comprehensive comparative analysis demonstrates that the TSNet model outperforms others, especially in the multi-point forecasting task. This study provides a valuable insight into the effectiveness of machine learning approaches and highlights the potential of the accuracy improvement for wave prediction.}, archive = {J_EAAI}, author = {Junhao Xu and Zhongying Feng and Zhan Wang and Kun Zheng and Ruipeng Li}, doi = {10.1016/j.engappai.2025.112385}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112385}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Twin-stream network: Enhancing wave prediction by capturing spatiotemporal information}, volume = {162}, year = {2025}, }
• (2025). Clean-sample guided cross-modal retrieval with adaptive weighted contrastive learning. EAAI, 162, 112383. (www)

  @article{ , abstract = {Cross-modal retrieval enables efficient integration of information by linking different data modalities, such as images and text. As data volumes increase rapidly, the need for effective cross-modal interaction grows. Cross-modal hashing is favored for its low storage requirements and fast retrieval speed, but many existing methods depend on accurately labeled data, which can be subjective and expensive to obtain. To address this limitation, we propose Clean-guided Adaptive Weighted Contrastive Hashing (CAWCH), a novel framework designed to improve robustness against noisy labels. CAWCH incorporates two main components: a Gaussian Mixture Model (GMM)-based noise purifier that identifies reliable and noisy samples by modeling sample loss, and a contrastive learning strategy that selectively chooses positive samples and adaptively assigns weights based on multi-label similarity, considering both intra- and inter-modal relationships. Extensive experiments demonstrate that CAWCH significantly outperforms existing methods under noisy label conditions, highlighting its effectiveness and potential for real-world cross-modal retrieval applications.}, archive = {J_EAAI}, author = {Shuni Jiang and Zhixin Li}, doi = {10.1016/j.engappai.2025.112383}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112383}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Clean-sample guided cross-modal retrieval with adaptive weighted contrastive learning}, volume = {162}, year = {2025}, }
• (2025). Denoising of the magnetic flux leakage signal using dynamic feature fusion and a multi-scale autoencoder network. EAAI, 162, 112382. (www)

  @article{ , abstract = {Magnetic flux leakage (MFL) signal denoising is essential for the nondestructive inspection of oil and gas pipelines, where complex noise interference can severely degrade defect quantification accuracy. Traditional approaches such as mean filtering and wavelet transform offer limited suppression of multi-type mixed noise and often distort critical features, including defect peaks and valleys. Even deep learning–based MFL denoising methods struggle in scenarios with substantial signal-noise overlap due to inadequate feature extraction and limited adaptability.This work presents an advanced denoising framework that combines dynamic feature fusion with a multi-scale autoencoder network. The framework jointly exploits time- and frequency-domain signal components, employing an adaptive weighting mechanism for dynamic feature fusion. Parallel convolutional branches extract multi-scale features, improving the capture of both global structures and fine-grained details, while a Squeeze-and-Excitation (SE) channel attention mechanism enhances defect-sensitive features and suppresses noise. Extensive experiments demonstrate that the proposed model outperforms mean filtering, wavelet denoising, and a baseline autoencoder, achieving notable gains in signal-to-noise ratio (SNR), mean squared error (MSE), and signal similarity. Beyond superior noise suppression, the method preserves critical defect characteristics, providing a robust and reliable foundation for precise defect quantification in pipeline MFL inspection.}, archive = {J_EAAI}, author = {Lushuai Xu and Shaohua Dong and Haotian Wei and Feng Li and Pengkun Zhang and Cong Zuo and Mingxing Guo and Penghui Liao}, doi = {10.1016/j.engappai.2025.112382}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112382}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Denoising of the magnetic flux leakage signal using dynamic feature fusion and a multi-scale autoencoder network}, volume = {162}, year = {2025}, }
• (2025). An optimisation approach guided by crack variation mechanism in the informer prediction model. EAAI, 162, 112381. (www)

  @article{ , abstract = {Structural health monitoring (SHM) faces a fundamental challenge in reconciling predictive performance with physical interpretability for infrastructure diagnostics. Conventional deep learning (DL) approaches neglect essential mechanisms governing crack width variation—including thermal gradients, hysteretic responses, and phase-shifted correlations—limiting their reliability in real-world applications. To bridge this gap, we propose a mechanism-guided optimization (MGO) framework that integrates domain knowledge into the Informer architecture through physics-informed enhancements: auto-correlation modeling for capturing temperature-crack hysteresis, static gated fusion for multi-feature integration, and adaptive elastic net regularization for feature selection. Validated on cable-stayed bridge monitoring data, our framework achieves significant mean absolute error reductions (MAE) (5 %–60 %) and root mean square error reductions (RMSE) (10 %–55 %) versus baseline Informer across all cracks and prediction horizons, with diebold-mariano (DM) tests confirming statistical superiority in most cases. Crucially, it demonstrates superior precision relative to six state-of-the-art benchmarks across all evaluation scenarios. The ordinary least squares (OLS)-enhanced variant further delivers volatility reduction, while sensor failure tests establish quantifiable robustness benchmarks through MAE progression from 0.013 mm to 0.391 mm. This work establishes an interpretable, physics-grounded paradigm that explicitly links environmental drivers to structural degradation.}, archive = {J_EAAI}, author = {Xujia Liu and Youliang Ding and Fei Xu and Yichao Xu and Kang Yang}, doi = {10.1016/j.engappai.2025.112381}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112381}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An optimisation approach guided by crack variation mechanism in the informer prediction model}, volume = {162}, year = {2025}, }
• (2025). A lightweight segmentation model based on segment anything model for tongue image segmentation. EAAI, 162, 112379. (www)

  @article{ , abstract = {Tongue image segmentation plays a crucial role in intelligent of diagnosis of Traditional Chinese Medicine. Accurate, efficient, and lightweight tongue segmentation significantly improves both the quality and practical applicability of intelligent disease diagnosis models. To address this challenge, we propose TongueSAM_Lite, a lightweight and fully automated tongue image segmentation model. Based on the Segment Anything Model. Our approach employs knowledge distillation and parameter-efficient fine-tuning to develop a novel lightweight image encoder, high-parameter modules in the Vision Transformer are partially replaced with lightweight image modules, which facilitate the transfer of its feature extraction capabilities while accelerating inference speed and reducing computational resource requirements. Additionally, to eliminate manual annotation of tongue region bounding boxes, we integrate a YOLOX-based automatic Box-prompt generator, enabling end-to-end fully automated prompting and segmentation of tongue images. To validate our approach, various experiments were conducted in three datasets. The results show that compared to the original large-scale model of the Segment Anything Model, TongueSAM_Lite reduces the size of the model by 42.7% and shortens the inference time to 45.43% while retaining the near-complete segmentation accuracy of few-shot learning. TongueSAM_Lite achieves Mean Intersection over Union scores of 96.48%, 98.36%, and 97.53% in the three datasets, respectively, outperforming state-of-the-art segmentation methods. Further validation confirms that the YOLOX-based prompt encoder yields optimal performance for the generation of tongue image bounding boxes. Our proposed approach provides new research insights to advance tongue diagnosis technology of Traditional Chinese Medicine. All codes in this article are available at https://github.com/ruanqunsheng/TongueSAM_Lite .}, archive = {J_EAAI}, author = {Qunsheng Ruan and Shan Cao and Zhirong Luo}, doi = {10.1016/j.engappai.2025.112379}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112379}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight segmentation model based on segment anything model for tongue image segmentation}, volume = {162}, year = {2025}, }
• (2025). Exploring cross-branch information for semi-supervised remote sensing object detection. EAAI, 162, 112378. (www)

  @article{ , abstract = {Semi-supervised object detection (SSOD) provides a promising solution to mitigate the annotation costs in remote sensing applications. Mainstream teacher-student based SSOD methods leverage unlabeled images through pseudo labeling, and their effectiveness is fundamentally limited by the inevitable noise in pseudo labels, particularly for remote sensing (RS) scenarios with complex backgrounds and dense, multi-scale and oriented objects. Current methods primarily focus on reducing pseudo label noise through category, scale and Intersection over Union information mining, as well as designing fine-grained confidence thresholding strategies. However, the inherent discrepancy between classification and localization reliability is neglected. In this study, with analyzing the characteristic discrepancies between the classification and localization branches, We propose artificial intelligence (AI) methodological innovation method named cross-branch information incorporation method (i.e., CBI-SSOD) to utilize these discrepancies to assist the training of the classification branch, and thus improve the performance of SSOD methods. Specifically, our method present two key AI innovations. Firstly, we propose a pretext task to extract cross-branch information, which can improve the classification ability by reinforce the consistent predictions between the classification branch and the pretext task. Besides, we propose a pseudo label reassignment approach to adjust the soft classification pseudo labels, and thus suppress pseudo label noise and improve the detection performance. Extensive experiments on Dataset for Object Detection in Aerial Images (DOTAv1.0) and DOTAv1.5 datasets validate the effectiveness and superiority of our method, and demonstrate the practical engineering impact of our method on RS applications and interpretation systems.}, archive = {J_EAAI}, author = {Shitian He and Huanxin Zou and Yingqian Wang and Xu Cao and Hao Chen and Ning Jing}, doi = {10.1016/j.engappai.2025.112378}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112378}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Exploring cross-branch information for semi-supervised remote sensing object detection}, volume = {162}, year = {2025}, }
• (2025). Lightning risk assessment of offshore wind farms considering tropical cyclone impacts based on an improved gradient boosting algorithm. EAAI, 162, 112376. (www)

  @article{ , abstract = {In recent years, global climate warming has led to a significant increase in both the frequency and intensity of tropical cyclones (TCs). The development of TCs is often accompanied by frequent lightning activities. The risk of lightning strikes to high offshore wind turbines is substantially elevated. This study evaluates the lightning risk faced by offshore wind farms influenced by tropical cyclones. Firstly, TC paths are analyzed in both spatial and temporal dimensions by linking them with lightning data to examine the distribution of TC-related lightning, and the lightning strike characteristics of offshore wind turbines are investigated. Secondly, a Bayesian Optimization (BO)-based eXtreme Gradient Boosting (XGBoost) model for lightning risk assessment is proposed, incorporating characteristics of TC lightning and offshore wind farms as input variables. The proposed BO-XGBoost model outperforms XGBoost, Bidirectional Long Short-Term Memory (Bi-LSTM), Support Vector Machine (SVM) and Neural Network (NN), achieving a precision of 98.9 % and a recall of 98.9 % on the test set. Additionally, SHapley Additive exPlanations (SHAP) value analysis indicates that TC lightning characteristics and offshore wind farm characteristics significantly impact the model output, enhancing the accuracy of the model. The assessment outcomes provide a theoretical basis for future offshore wind farm planning and guidance for lightning protection measures in offshore wind farms.}, archive = {J_EAAI}, author = {Qibin Zhou and Kehan Chen and Xiaoyan Bian and Shangjie Chen and Gaopeng Lu}, doi = {10.1016/j.engappai.2025.112376}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112376}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Lightning risk assessment of offshore wind farms considering tropical cyclone impacts based on an improved gradient boosting algorithm}, volume = {162}, year = {2025}, }
• (2025). Beyond trial-and-error: Predicting user abandonment after a moderation intervention. EAAI, 162, 112375. (www)

  @article{ , abstract = {Current content moderation follows a reactive, trial-and-error approach, where interventions are applied and their effects are only measured post-hoc. In contrast, we introduce a proactive, predictive approach that enables moderators to anticipate the impact of their actions before implementation. We propose and tackle the new task of predicting user abandonment following a moderation intervention. We study the reactions of 16,540 users to a massive ban of online communities on Reddit, training a set of binary classifiers to identify those users who would abandon the platform after the intervention—a problem of great practical relevance. We leverage a dataset of 13.8 million posts to compute a large and diverse set of 142 features, which convey information about the activity, toxicity, relations, and writing style of the users. We obtain promising results, with the best-performing model achieving micro F1-score = 0 . 914 . Our model shows robust generalizability when applied to users from previously unseen communities. Furthermore, we identify activity features as the most informative predictors, followed by relational and toxicity features, while writing style features exhibit limited utility. Theoretically, our results demonstrate the feasibility of adopting a predictive machine learning approach to estimate the effects of moderation interventions. Practically, this work marks a fundamental shift from reactive to predictive moderation, equipping platform administrators with intelligent tools to strategically plan interventions, minimize unintended consequences, and optimize user engagement.}, archive = {J_EAAI}, author = {Benedetta Tessa and Lorenzo Cima and Amaury Trujillo and Marco Avvenuti and Stefano Cresci}, doi = {10.1016/j.engappai.2025.112375}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112375}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Beyond trial-and-error: Predicting user abandonment after a moderation intervention}, volume = {162}, year = {2025}, }
• (2025). A sampling interval-adaptive transformer for industrial time sequence modeling with heterogeneou s sampling rates in quality prediction. EAAI, 162, 112374. (www)

  @article{ , abstract = {The industrial data sequences frequently exhibit irregular sampling frequencies, which pose a number of difficulties for data analysis and modeling. The traditional dynamic models like Recurrent Neural Network (RNN) and Transformer are difficult to model such data sequences. The main reason is that these models assume that data sampling frequency should be constant. To this end, a Sampling Interval-Adaptive Transformer (SIA-Trans) is proposed in this paper to adaptively model the temporal information for heterogeneous sampling sequences in industrial processes. The SIA-Trans uses the sampling interval and position embedding block to address the problem of unequal time intervals and rectify the temporal correlations in time series. Then, the interval-aware self-attention net is designed for dynamic data relationship modeling, taking the processed data through the self-attention mechanism. Finally, the predicted output is obtained after the point-wise feed-forward layer. The proposed SIA-Trans is validated on a real-world hydrocracking process to predict the content of hydrocarbon mixture with five carbon atoms (C5) hydrocarbons in light naphtha, as well as the final boiling point of jet fuel.}, archive = {J_EAAI}, author = {Zijian Xu and Nuo Xu and Kai Wang and Xiaofeng Yuan and Yalin Wang and Chunhua Yang and Weihua Gui and Shuqiao Cheng and Lingjian Ye}, doi = {10.1016/j.engappai.2025.112374}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112374}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A sampling interval-adaptive transformer for industrial time sequence modeling with heterogeneou s sampling rates in quality prediction}, volume = {162}, year = {2025}, }
• (2025). Multi-modality complementary learning network with cross-modality interaction and adaptive fusion for face forgery detection. EAAI, 162, 112373. (www)

  @article{ , abstract = {With the advancement of face forgery technology, many forged faces threaten information security. Although existing face forgery detection methods obtain better detection performance on intra-dataset evaluation, the generalization of cross-dataset detection and the robustness against image post-processing operations still need to be improved. To address these issues, we propose a multi-modality complementary learning network with cross-modality interaction and adaptive fusion for face forgery detection. Specifically, a two-branch architecture is designed to extract spatial and frequency features. To realize the interaction and communication of spatial and frequency information, the cross-modality interaction module is designed to explore the inter-modality correlation by applying across self-attention. Subsequently, a multi-scale feature enhancement module is introduced in the spatial branch to enhance the texture and semantic information of spatial features, improving the robustness of tackling image post-processing operations. In addition, to exploit the complementary relationship between the spatial and frequency features, an adaptive fusion module is designed to establish forged feature dependencies by leveraging spatial self-attention, while learning discriminative feature representations by fusing spatial and frequency features in an adaptive weighted manner. Extensive experimental results on four public datasets demonstrate that the proposed method outperforms other state-of-the-art methods for intra-dataset, cross-dataset, and perturbed dataset evaluations.}, archive = {J_EAAI}, author = {Chunyin Shi and Chengyou Wang and Xiao Zhou and Zhiliang Qin}, doi = {10.1016/j.engappai.2025.112373}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112373}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-modality complementary learning network with cross-modality interaction and adaptive fusion for face forgery detection}, volume = {162}, year = {2025}, }
• (2025). Adaptive neural network tracking control for unknown high-order nonlinear systems: A constructive approximation set based approach. EAAI, 162, 112371. (www)

  @article{ , abstract = {This article addresses the problem of adaptive neural network (NN) tracking control for unknown high-order nonlinear systems, with a focus on accurately constructing NN approximation sets. To guarantee the local approximation capabilities of NNs, it is crucial that their input signals remain within corresponding compact sets. However, the unknown functions and powers in high-order nonlinear systems make it difficult to determine these sets accurately. To solve this, we introduce a novel adaptive NN tracking control strategy that integrates signal substitution technique, barrier functions (BFs), and NNs. Specifically, the signal substitution technique converts the original system states into state error variables, along with the desired reference signal and its time derivatives, which serve as part of the NN input. BFs are employed to constrain the state errors, while NNs approximate the transformed unknown system functions. This approach enables precise calculation of bounds for the NN weight estimators, ensuring that the NN approximation sets are constructed. Unlike existing methods, our approach not only proves the existence of NN approximation sets but also provides a constructive design strategy, significantly enhancing the approximation accuracy for unknown nonlinear functions. Simulation results demonstrate the effectiveness and advantages of the proposed method.}, archive = {J_EAAI}, author = {Yu-Fa Liu and Yong-Hua Liu and Jin-Wa Wu and Jie Tao and Ming Lin and Chun-Yi Su and Renquan Lu}, doi = {10.1016/j.engappai.2025.112371}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112371}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive neural network tracking control for unknown high-order nonlinear systems: A constructive approximation set based approach}, volume = {162}, year = {2025}, }
• (2025). Dynamic correlation graph convolution network with embedded temporal correlation extraction for stock price forecasting. EAAI, 162, 112370. (www)

  @article{ , abstract = {Stock price forecasting has become a significant and complex research area within financial technology. The dynamic correlations among stocks and the inherent noise in price volatility present considerable challenges in accurately forecasting stock prices and enhancing investment returns. This paper introduces a novel Dynamic Correlation Graph Convolution Network (DyCGCN) with embedded temporal correlation extraction. First, we propose a dual-scale dynamic graph generation method to capture the topological relationships among stocks. Second, we develop a dynamic correlation-temporal convolution module that extracts high-level temporal correlations. Third, we introduce a prospect theory-guided multi-strategy loss function that accommodates the diverse risk preferences of investors. Furthermore, we present a joint regression-classification learning method to extract and leverage stock trend information. Experiments conducted on four real-world datasets demonstrate the superiority of DyCGCN, achieving an average 24.7% reduction in prediction error and a 10.5% improvement in predictive accuracy over baseline models, underscoring its strong potential for practical stock price forecasting.}, archive = {J_EAAI}, author = {Fang He and Wei Yin and Yilun Jin and Zhengyang Chen}, doi = {10.1016/j.engappai.2025.112370}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112370}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dynamic correlation graph convolution network with embedded temporal correlation extraction for stock price forecasting}, volume = {162}, year = {2025}, }
• (2025). Multi-objective deployment optimization for integrated sensing and communication-enabled unmanned aerial vehicle swarm. EAAI, 162, 112368. (www)

  @article{ , abstract = {With the convergence of mobile communication, sensing, and computational networks in sixth-generation technology, the integration of sensing and communication with unmanned aerial vehicles (UAVs) is promising. This paper focuses on the contribution of artificial intelligence in optimizing the deployment of UAV swarms for multi-objective target detection applications in sixth-generation networks. Specifically, the artificial intelligence contribution lies in the development of an improved multi-objective particle swarm optimization (IMOPSO) algorithm for solving a complex multi-objective deployment problem. The problem aims to simultaneously optimize communication rate, sensing quality, and energy consumption in the deployment of UAV swarms. To address this, the proposed IMOPSO incorporates chaotic initialization, Lévy flight mutation, dynamic mutation rate, and an elimination mechanism based on opposition-based learning. These innovations are designed to enhance the algorithm’s ability to explore the solution space effectively, overcome premature convergence to local solutions, and improve solution quality. In terms of engineering applications, the IMOPSO is applied to the deployment of UAV swarms for target detection, demonstrating its ability to enhance communication and sensing performance while reducing energy consumption in practical scenarios. Through extensive simulations, we show that the IMOPSO outperforms traditional optimization methods and other baseline algorithms, achieving superior results across all optimization objectives. Specifically, the IMOPSO achieves approximately 5% higher transmission data rate, 9% better sensing quality, and 19% lower energy consumption compared to baseline algorithms across multiple test scenarios. Furthermore, the solutions obtained are not only closer to the optimal front but also more concentrated, indicating higher-quality results.}, archive = {J_EAAI}, author = {Hongjuan Li and Haiyuan Chen and Miao Wang and Jiahui Li and Hui Kang and Yuzhuo Guan and Xu Lin}, doi = {10.1016/j.engappai.2025.112368}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112368}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-objective deployment optimization for integrated sensing and communication-enabled unmanned aerial vehicle swarm}, volume = {162}, year = {2025}, }
• (2025). Progressive semi-supervised learning for specific emitter identification: An iterative clustering and pseudo-label refinement approach. EAAI, 162, 112361. (www)

  @article{ , abstract = {Specific Emitter Identification (SEI) distinguishes radio-frequency (RF) devices by exploiting hardware-induced signal fingerprints, thereby strengthening wireless-layer security. Existing deep learning-based SEI methods depend heavily on labeled data and fixed confidence thresholds for pseudo-labeling, which limits their effectiveness under label scarcity or open-set conditions. To overcome these issues, we propose a progressive semi-supervised learning (ProSSL) method for SEI that combines iterative clustering with contrastive learning to generate adaptive pseudo-labels. ProSSL introduces an “uncertain” class and employs a dual-constraint selector—prediction stability and class diversity—to suppress noisy pseudo-labels and ensure robust propagation. Experiments on the public real-world long range(LoRa) RF-fingerprint dataset show that ProSSL gains 2.90%–6.01% absolute accuracy over state-of-the-art baselines, reaching 96.48% accuracy with 90% labels and 59.88% with only 5% labels. While on the public automatic dependent surveillance-broadcast(ADS-B) Top-10 dataset, ProSSL achieves 84.40% accuracy with 5% labeled data and 99.40% accuracy with 90%labeled data, again outperforming all competing methods. Open-set evaluations further demonstrate that overall accuracy rises from 18.81% when only two classes are known to 62.25% when eight classes are known, confirming strong generalization to unseen emitters and validating ProSSL’s robustness and practicality in realistic wireless environments.}, archive = {J_EAAI}, author = {Yiting Gao and Ke Wang and Hao Huang and Jiao Wang and Jiaxu Liu and Yao Zheng and Jianqing Li}, doi = {10.1016/j.engappai.2025.112361}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112361}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Progressive semi-supervised learning for specific emitter identification: An iterative clustering and pseudo-label refinement approach}, volume = {162}, year = {2025}, }
• (2025). Enhancing scene text image super-resolution via gradient-based graph attention network. EAAI, 162, 112360. (www)

  @article{ , abstract = {Scene text image super-resolution is crucial for enhancing text recognition in low-resolution real-world images. Existing methods usually overlook the structured and repetitive layout of text, which can serve as powerful prior knowledge for guiding reconstruction. In this work, we propose a novel framework that incorporates gradient-based graph attention to explicitly model patch-level text layout. The architecture combines a non-local group-wise attention module, a cascaded channel attention module, and a gradient-guided graph attention module to capture both global and local structural dependencies. This design enables more accurate restoration of text contours and layout consistency. Extensive experiments on the benchmark dataset demonstrate that our method achieves superior performance in both image quality and recognition accuracy, outperforming state-of-the-art methods. The code is available at: https://github.com/cvzxy/TSANv2 .}, archive = {J_EAAI}, author = {Xiangyuan Zhu and Xuchong Liu and Kehua Guo and Wei Zhao}, doi = {10.1016/j.engappai.2025.112360}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112360}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing scene text image super-resolution via gradient-based graph attention network}, volume = {162}, year = {2025}, }
• (2025). A consistency regularization-based approach integrating anatomical structural relationships and organ category representations for multi-organ segmentation in pigs. EAAI, 162, 112359. (www)

  @article{ , abstract = {In modern biomedical research and livestock management, accurate multi-organ segmentation in pigs is essential for breeding programs. However, current methods face challenges due to low imaging contrast, size disparities, and organ shape variability. Additionally, the manual annotation of computed tomography (CT) scans is labor-intensive and costly, limiting available labeled samples. To address these issues, we propose a consistency regularization-based network guided by anatomical structural relationships and global organ category representations, specifically designed for multi-organ segmentation using a limited number of annotated CT scan samples from pigs. Specifically, we designed the SpatialLink Gated Recurrent Unit (GRU) module to extract anatomical structural information and capture dynamic spatial relationships between organs, thereby minimizing segmentation biases caused by organ shape variations. Moreover, we developed the Organ Category Coding module and Guidance module, which integrate consistency regularization and attention mechanisms, enabling the network to accurately extract global organ category representations during the decoding phase, even with a small number of labeled samples, significantly improving segmentation consistency across organs of different sizes. Additionally, We are the first to apply the Visual State Space block to multi-organ segmentation in pigs, using it to extract contextual information. Experiments on 60 pigs demonstrate that our method achieves state-of-the-art results, with significant improvements in segmentation accuracy for the gallbladder and bladder, including a 9.8% and 4.2% Dice score increase, respectively, and a 12.4% and 6.2% boost in Jaccard scores compared to compared with a selection of published methods.}, archive = {J_EAAI}, author = {Xiang Pan and Hang Fan and Jianlan Wang and Yan Fu and Wei Chu and Weipeng Tai and Jing Gu and Jianming Ni}, doi = {10.1016/j.engappai.2025.112359}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112359}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A consistency regularization-based approach integrating anatomical structural relationships and organ category representations for multi-organ segmentation in pigs}, volume = {162}, year = {2025}, }
• (2025). Interpretable chest X-ray localization using principal component-based feature selection in deep learning. EAAI, 162, 112358. (www)

  @article{ , abstract = {Accurate identification and localization of diseases in chest X-ray (CXR) images are crucial for early diagnosis and timely medical intervention. Traditional localization techniques like Class Activation Mapping (CAM), depend on Global Average Pooling (GAP) layers, restricting their flexibility, while gradient-based methods like Grad-CAM involve computational overhead and limited interpretability. To address these limitations, this study introduces a novel Principal Component Analysis (PCA)-based localization method that eliminates reliance on GAP layers and gradient computations. Utilizing publicly available Kaggle datasets, namely the COVID-19 Radiography Dataset and Tuberculosis (TB) Chest X-ray Database. The proposed approach employs PCA to compress high-dimensional convolutional feature maps extracted from the pretrained VGG16 model into a lower-dimensional, spatially meaningful representation. This enables rapid, interpretable heatmap generation highlighting precise abnormal regions. Experimental results demonstrate that the proposed method achieved an average training loss of 0 . 0835 ± 0 . 1830 and validation loss of 0 . 1385 ± 0 . 0741 across 5-fold cross-validation. In addition, it achieved an impressive accuracy of 97.5%, sensitivity of 98.2%, specificity of 99.4%, a Dice Similarity Coefficient (DSC) of 97.5%, and an Intersection-over-Union (IoU) of 95.1%. Compared to CAM, and Grad-CAM, PCA-based localization significantly reduces inference time, enhances interpretability, and provides robust multi-class localization performance suitable for clinical deployment.}, archive = {J_EAAI}, author = {Diwakar Diwakar and Deepa Raj}, doi = {10.1016/j.engappai.2025.112358}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112358}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Interpretable chest X-ray localization using principal component-based feature selection in deep learning}, volume = {162}, year = {2025}, }
• (2025). Fast simulation for scattering muography applications using generative adversarial neural networks. EAAI, 162, 112357. (www)

  @article{ , abstract = {Muography is an emergent non-destructive testing technique that uses cosmic muons to probe the interior of objects and structures. This technique can be employed to perform preventive maintenance of critical equipment in the industry in order to test the structural integrity of the facility. Several muography imaging algorithms based on machine learning methods are being developed in the recent years. These algorithms make exhaustive use of simulated data, usually using packages such as GEANT4 (GEometry ANd Tracking), that exhaustively simulate the detector, to produce training samples. This work presents a faster alternative for the generation of simulated samples based on generative adversarial neural networks. A speed up factor of 80 is observed with this system without any significant degradation of the quality of the simulation.}, archive = {J_EAAI}, author = {Rubén López Ruiz and Celia Fernández Madrazo and Sergio Sánchez Cruz and Lara Lloret Iglesias and Pablo Martínez Ruiz del Árbol}, doi = {10.1016/j.engappai.2025.112357}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112357}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fast simulation for scattering muography applications using generative adversarial neural networks}, volume = {162}, year = {2025}, }
• (2025). Research on lightweight network for real-time detection of steel structure corrosion based on knowledge distillation. EAAI, 162, 112356. (www)

  @article{ , abstract = {Regular and real-time monitoring of corrosion is crucial for ensuring structural safety and extending the service life of infrastructure. With the continuous development of advanced structural health monitoring technologies, intelligent corrosion detection has become an inevitable trend. This study addresses the issues of low accuracy, incomplete detail processing, and missed or false detections in complex scenarios in steel structure corrosion detection, as well as the challenges of high complexity and insufficient real-time performance in deep learning models. We propose a high-performance, lightweight, real-time corrosion detection model, Real-Time Detection Transformer for Corrosion (RT-DETR-Corrosion), based on knowledge distillation. By incorporating lightweight optimization on the RT-DETR-R18 baseline model and a hybrid knowledge distillation approach, the model significantly improves real-time performance and detection accuracy, meeting the application requirements for efficiency and precision in steel structure corrosion detection. Experimental results show that the model exhibits excellent optimization effects in terms of localization accuracy, classification accuracy, and position regression error on both the training and validation sets, while also demonstrating strong generalization ability. In extreme weather conditions (such as rain, fog, snow, and strong light) and complex scenarios (such as occlusion, blur, and low-light environments), the model maintains stable Precision, Recall, and mAP metrics, validating its reliability and applicability in diverse real-world engineering environments. Moreover, visualized heatmap analysis of detection results for different scenarios further confirms the model's precise attention to corrosion regions and its generalization ability, providing essential technical support for steel structure corrosion risk assessment and intelligent monitoring, with significant potential for engineering applications.}, archive = {J_EAAI}, author = {Jia Hou and Wei Chen and Zhen Duan and Hang Li and Mingyu Yu}, doi = {10.1016/j.engappai.2025.112356}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112356}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Research on lightweight network for real-time detection of steel structure corrosion based on knowledge distillation}, volume = {162}, year = {2025}, }
• (2025). Function structures of information measures for n-polygonal interval-valued intuitionistic fuzzy sets and their application in location selection strategy. EAAI, 162, 112355. (www)

  @article{ , abstract = {The interval-valued intuitionistic fuzzy set (IVIFS) represents an organic integration of interval-valued fuzzy sets and intuitionistic fuzzy sets, but it fails to meet the requirements of linearity and closure in arithmetic operations, rendering its computations relatively intricate. Consequently, this paper employs an interval partitioning strategy to mitigate the complexity of arithmetic operations, thereby constructing a novel fuzzy set structure characterized by efficient piecewise linear approximation capabilities. Furthermore, the paper presents the structural form of arithmetic operations for the IVIFS with piecewise linear approximation and demonstrates the simplicity of these operations by an numerical example. In addition, we extend the findings on information measures for polygonal interval-valued intuitionistic fuzzy set pertaining to abstract functions that fulfill particular criteria. Furthermore, we delve into the transformation relationships among these information measures, from which a range of structural formats for information measures can be deduced based on functional expressions and transformation relationships. Finally, similarity measures are applied to company site selection. The validity, practicality and stability of proposed measures have been proven through sensitivity analysis and comparative analysis.}, archive = {J_EAAI}, author = {Le Fu and Chunfeng Suo}, doi = {10.1016/j.engappai.2025.112355}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112355}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Function structures of information measures for n-polygonal interval-valued intuitionistic fuzzy sets and their application in location selection strategy}, volume = {162}, year = {2025}, }
• (2025). Two-dimensional refined composite multi-scale revised ensemble dispersion entropy and its application to fault diagnosis of rolling bearing. EAAI, 162, 112354. (www)

  @article{ , abstract = {One-dimensional ensemble dispersion entropy (EDE 1D ) can effectively characterize the nonlinear dynamic characteristics of one-dimensional time series, but the complexity of two-dimensional space is not reflected, and only single-scale features can be captured. Firstly, to comprehensively capture the feature information of two-dimensional space, the symmetrized dot pattern (SDP) is introduced to overcome the shortcomings of the ordinary images lack of physical meaning and the time-frequency distribution methods exhibit incomplete information representation, etc. Simultaneously, the amplitude and frequency information are intuitively expressed by a two-dimensional mirror snowflake symmetrized image (MSSI 2D ). Secondly, to overcome the shortcomings of single-scale and traditional coarse-graining, a two-dimensional refined composite multi-scale coarse-graining method is proposed, which improves the accuracy of feature extraction and reduces the calculation deviation. After that, a new feature extraction method namely two-dimensional refined composite multi-scale revised ensemble dispersion entropy (RCMREDE 2D ) is proposed, whose parameter stability and performance are explored through simulation analysis. The results demonstrate that the RCMREDE 2D exhibits excellent stability and anti-noise interference ability. Based on the advantages of RCMREDE 2D , a novel fault diagnosis method for rolling bearings is developed by integrating RCMREDE 2D and a firefly algorithm optimized support vector machine (FA-SVM) multi-fault classifier for pattern recognition. The proposed method is further validated through two measured bearing data sets and five comparative methods, and the results indicate that the RCMREDE 2D and FA-SVM achieve the highest recognition accuracy while demonstrating superior stability.}, archive = {J_EAAI}, author = {Wenqing Ding and Jinde Zheng and Haiyang Pan and Jian Cheng and Jinyu Tong}, doi = {10.1016/j.engappai.2025.112354}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112354}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Two-dimensional refined composite multi-scale revised ensemble dispersion entropy and its application to fault diagnosis of rolling bearing}, volume = {162}, year = {2025}, }
• (2025). Sparse and robust elastic net support vector machine with bounded concave loss for large-scale problems. EAAI, 162, 112352. (www)

  @article{ , abstract = {The elastic net support vector machine is an extensively employed method for addressing a range of classification tasks. Nevertheless, a significant drawback of the elastic net support vector machine is its high computational cost when dealing with large-scale classification problems. To address this drawback, we first introduce an innovative non-convex elastic net support vector machine model that employs our newly created bounded concave loss function, which effectively attains both sparsity and robustness. Based on proximal stationary point, we have effectively constructed an innovative optimality theory tailored for our newly created elastic net support vector machine model. By leveraging the innovative optimality theory, we have successfully developed a new and exceptionally effective algorithm designed to enhance computational efficiency through the division of the entire dataset into two distinct categories: working sets and non-working sets. During each learning cycle, the parameters associated with the non-working set remain unchanged. In contrast, the parameters related to the working set are subject to updates. Consequently, our new algorithm facilitates quicker modifications on smaller datasets, improving runtime efficiency and lowering computational complexity. Numerical experiments have demonstrated significant efficiency, particularly regarding computational speed, the number of support vectors, and classification accuracy, surpassing eleven other leading solvers.}, archive = {J_EAAI}, author = {Huajun Wang and Wenqian Li}, doi = {10.1016/j.engappai.2025.112352}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112352}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Sparse and robust elastic net support vector machine with bounded concave loss for large-scale problems}, volume = {162}, year = {2025}, }
• (2025). A review of speaker verification: Methods, network architectures, tasks and challenges. EAAI, 162, 112351. (www)

  @article{ , abstract = {Speaker verification is an important branch of biometric recognition, with wide applications in identity authentication, audio monitoring, and other fields. In recent years, deep learning and meta-learning have made remarkable advancements in the field of speaker verification. Therefore, it is necessary to update existing reviews of speaker verification to reflect the latest research developments. We review literature from the past decade to provide a timely and comprehensive survey of the field. First, we outline the concept and system process of speaker verification. Then, we analyze the speech preprocessing process and common acoustic features used in the systems. Next, we present an overview of speaker modeling approaches, covering traditional probabilistic methods, deep learning-based speaker methods, and meta-learning-based speaker methods, focusing on the latter two methods. We provide an in-depth analysis and summary of the characteristics and the latest network architectures of these methods, focusing on the development of Transformer and large-scale pre-trained Transformer. Furthermore, we introduce the datasets and evaluation metrics used in speaker verification systems, focusing on a detailed and fair comparison of the performance of text-dependent and text-independent speaker verification systems. Finally, we explore the challenges faced by speaker verification systems and discuss future research opportunities.}, archive = {J_EAAI}, author = {Weijie Wang and Hong Zhao and Yikun Yang and Yongjuan Yang}, doi = {10.1016/j.engappai.2025.112351}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112351}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A review of speaker verification: Methods, network architectures, tasks and challenges}, volume = {162}, year = {2025}, }
• (2025). Low-cost and sparsity for continual semantic segmentation. EAAI, 162, 112350. (www)

  @article{ , abstract = {Deep neural networks have contributed to significant progress in semantic segmentation tasks. However, deep neural networks exhibit a critical drop in performance due to catastrophic forgetting when they are required to learn new tasks incrementally. The more plastic the network is, the easier it can learn new tasks. Whereas, for continual semantic segmentation, it is more reliable to preserve the knowledge it has learned from previous tasks. Here, gated 0-1 Bernoulli variable is used as a regularization method to optimize performance by enhancing network sparsity. Then, the special case of gated 0-1 Bernoulli variable is applied in the replay-based method of continual semantic segmentation. Specifically, when the value of the sub-network sampling rate reaches 0.5, the network reaches the strongest stability. Finally, the gated 0-1 Bernoulli variable improves the network’s performance in complex scenarios and reduces cost under similar performance. Experimental results indicate that in using 100% samples for incremental training, the Mean Intersection over Union(mIoU) of the old classes improves by up to 4.6% and 5.5% compared to the baseline at the end of the overall training in continual semantic segmentation scenarios 10-1 and 10-2. Furthermore, in using 60% samples for incremental training, the performance for the old tasks only drops by less than a percentage, while the time cost to complete the full setup decreases by 22%.}, archive = {J_EAAI}, author = {Qing Ji and Bin Li and Shaobo Li and Hongchao An and Jing Yang}, doi = {10.1016/j.engappai.2025.112350}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112350}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Low-cost and sparsity for continual semantic segmentation}, volume = {162}, year = {2025}, }
• (2025). Promoting sustainability-oriented brand activist campaigns: A spherical fuzzy decision support framework for evaluating activist advertising videos. EAAI, 162, 112349. (www)

  @article{ , abstract = {Activist video advertisements represent a strategic form of brand communication in which companies express their stance on social or environmental issues through emotionally driven storytelling and slogan-based narratives. Despite their growing prevalence, there is a notable lack of systematic and quantitative methods for evaluating their performance or comparing their effectiveness across competing brands. This research addresses that gap by proposing a comprehensive decision support system (DSS) designed to assess the performance of activist video advertisements in a structured and reproducible manner. The study introduces a novel hybrid multi-criteria decision-making (MCDM) framework: the spherical cubic fuzzy (SCF)–Aczel-Alsina–ranking comparison (RANCOM)–method based on the removal effects of criteria (MEREC)–deviation-based pairwise assessment ratio technique (DEPART). This methodology integrates subjective weights obtained via SCF–RANCOM and objective weights derived through SCF–MEREC, with both sets of weights combined using SCF-based aggregation operators that incorporate Aczel-Alsina t-norm and t-conorm functions. Performance rankings are then generated using the SCF–DEPART method. To demonstrate the model's applicability, a real-world case study involving eight sustainability-oriented activist video advertisements released in Türkiye was conducted. Evaluations were based on input from ten domain experts across eleven criteria. The analysis identified “convincingness and credibility” as the most critical factor, with “The Voice of Nature” campaign achieving the highest performance rating. The model's robustness was confirmed through scenario-based sensitivity analyses, and its consistency was validated by benchmarking against thirteen alternative MCDM approaches. The findings offer meaningful implications for both academic research and advertising practice.}, archive = {J_EAAI}, author = {Galip Cihan Yalçın and Karahan Kara and Gülcan Işık and Esra Serdar Tekeli and Vladimir Simic and Abdullah Ballı and Dragan Pamucar}, doi = {10.1016/j.engappai.2025.112349}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112349}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Promoting sustainability-oriented brand activist campaigns: A spherical fuzzy decision support framework for evaluating activist advertising videos}, volume = {162}, year = {2025}, }
• (2025). Federated clustering with mutual knowledge distillation for traffic flow prediction. EAAI, 162, 112347. (www)

  @article{ , abstract = {Traffic flow prediction plays a critical role in intelligent transportation systems. Conventional traffic flow prediction methods primarily rely on centralized training, which poses a risk of privacy leakage. Federated learning, a privacy-preserving framework to machine learning, enables distributed participants to jointly train a shared model without sharing local private data. However, traffic flow is typically collected from different devices and contains different temporal patterns, leading to non-independent and identically distributed. To address these challenges, we propose a traffic flow prediction method based on federated clustering with mutual knowledge distillation. We first perform temporal decomposition on the traffic flow data and use mutual learning with adaptive distillation loss to facilitate mutual knowledge transfer among local models during training. Then, we apply spectral clustering to cluster clients based on the cosine similarity of model parameters at the server and design a global model aggregation method to improve the performance of federated learning. Finally, the proposed method is evaluated on two real-world traffic datasets, and the experiment results show significant improvements over traditional federated learning approaches and also outperform federated mutual learning. The results demonstrate that the proposed method effectively captures temporal information and mitigates the effect of non-independent and identically distributed issues.}, archive = {J_EAAI}, author = {Yao Lin and Shengwu Xiong}, doi = {10.1016/j.engappai.2025.112347}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112347}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Federated clustering with mutual knowledge distillation for traffic flow prediction}, volume = {162}, year = {2025}, }
• (2025). Quantifying hybrid failure mode in cyclic-loaded reinforced concrete shear walls: Integrating unsupervised and supervised learning techniques. EAAI, 162, 112346. (www)

  @article{ , abstract = {This paper aims to identify failure types in Reinforced Concrete Shear Walls (RCSWs) by quantifying the contributions of shear and flexural modes in the force-deformation response of the walls. The supporting research database includes images, cyclic curve backbones, and geometric and mechanical characteristics of 253 RCSWs. Initially, the database is manually classified into shear, flexure, and shear-flexure failure modes based on observations of surface damage and the cyclic response of the wall. Subsequently, unsupervised clustering and supervised learning algorithms are employed to probabilistically quantify and predict the participation of flexural and shear modes in the overall seismic response of the walls, respectively. The unsupervised model, utilizing the K-means algorithm, identifies the primary failure modes of the walls, achieving over 90 % concordance with manual expert labeling. Based on the results of unsupervised clustering, a hybridity index is proposed to demonstrate the contributions of shear and flexure failure modes to the overall seismic response. Supervised learning is then used to predict hybridity indices from wall characteristics, with the Extremely Randomized Trees (Extra Trees) model achieving the best results based on a balanced evaluation of multiple performance metrics. SHapley Additive exPlanations (SHAP), a tool for exploring model sensitivity, highlights the aspect ratio as the key influencing factor on failure mode, in accordance with relevant structural engineering codes and standards. Implementation of the proposed framework in exploring the behaviors of four unseen case studies reveals a significant correlation between the predicted hybridity indices and observed damage, consistent with existing guidelines.}, archive = {J_EAAI}, author = {Pouya Ebrahimi and Amir Hossein Asjodi and Kiarash M. Dolatshahi}, doi = {10.1016/j.engappai.2025.112346}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112346}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Quantifying hybrid failure mode in cyclic-loaded reinforced concrete shear walls: Integrating unsupervised and supervised learning techniques}, volume = {162}, year = {2025}, }
• (2025). Dual-path multiple attention-guided feature interaction network for camouflaged object detection. EAAI, 162, 112345. (www)

  @article{ , abstract = {Existing Camouflaged Object Detection (COD) methods rely on features from a pre-trained backbone with a single-encoder structure, leading to initial features biased towards global or local preferences. This affects subsequent feature modeling and degrades COD performance. Some studies use modular or auxiliary flow structures to balance feature preferences but often focus only on interactions between hierarchical features or information flows, ignoring potential redundancy or noise within aggregated features. Therefore, we propose a novel dual-path multiple attention-guided feature interaction network (DMAFI-Net) for COD, which contains four main components: global and local features interaction module (GLFI), intra-feature interaction module (IFI), multi-scale feature enhancement module (MFE), and two decoders including neighbor connection decoder based feature aggregation (NFA) and refine decoder. Specifically, the GLFI is designed to implement the interaction and combination of global and local features, and the combined features will be sent to IFI to mine intra-feature information. Besides, the MFE is introduced to further enrich the extracted features obtained in the IFI. In the feature decoding stage, the NFA module utilizes neighbor connection decoder to fuse multi-scale features and ultimately generates a coarse prediction. Finally, the refine decoder leverages multiple attention modules to refine the initial prediction with the combined features as auxiliary cues and obtain the final camouflaged map. Extensive experiments on four COD benchmark datasets demonstrate the superiority of the proposed framework when compared to 24 state-of-the-art (SOTA) methods in terms of five widely used evaluation metrics. Furthermore, the ablation studies show the effectiveness of main components of our DMAFI-Net.}, archive = {J_EAAI}, author = {Anzhi Wang and Jintao Wu and Shuang Zhao and Yun Liu}, doi = {10.1016/j.engappai.2025.112345}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112345}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dual-path multiple attention-guided feature interaction network for camouflaged object detection}, volume = {162}, year = {2025}, }
• (2025). TFCTL: Time–Frequency calibrated transfer learning for cross domain depression detection. EAAI, 162, 112344. (www)

  @article{ , abstract = {As the application of speech-based depression detection systems expands, differences in cross-domain data distribution pose significant challenges. This paper proposes the framework of Time–Frequency Calibrated Transfer Learning (TFCTL). This framework first introduces Frequency-Delay Neural Network (FDNN), inspired by Time-Delay Neural Network (TDNN), extending the concept of temporal feature extraction using sliding windows and weight sharing from the time domain to the frequency domain. A multi-level information aggregation module then integrates features of varying abstraction levels from both time-delay and frequency-delay neural networks, balancing global and local speech information. Finally, TFCTL uses transfer learning to calibrate the distribution of the aggregated time–frequency embedding vectors, uncovering commonalities of depression features across different domains. Cross-speaker and cross-corpus experiments were conducted using the Chinese Multimodal Depression Corpus (CMDC) and the Distress Analysis Interview Corpus Wizard-of-Oz (DAIC-WOZ). In cross-speaker scenarios, TFCTL achieved F1 scores of 0.7324 on DAIC and 0.9660 on CMDC, outperforming other methods. In cross-corpus scenarios, TFCTL achieved F1 scores of 0.6743 on DAIC and 0.6879 on CMDC, demonstrating its robustness in addressing domain mismatch issues. The source code used in the paper is available at https://anonymous.4open.science/r/TFCTL-A545/ .}, archive = {J_EAAI}, author = {Dongdong Li and Li Ding and Zuo Yang and Zhe Wang and Ke Zhao}, doi = {10.1016/j.engappai.2025.112344}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112344}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {TFCTL: Time–Frequency calibrated transfer learning for cross domain depression detection}, volume = {162}, year = {2025}, }
• (2025). Mutual information guided invertible image hiding network. EAAI, 162, 112343. (www)

  @article{ , abstract = {Image hiding techniques are commonly used for secure communication, copyright protection, and visual privacy. Invertible neural network (INN) have emerged as a promising approach for image steganography, enabling the concealment and recovery of secret images through forward and backward mappings within the network. However, existing methods often face limitations in the accuracy of recovered images due to challenges in estimating the lost information during the forward process. To address this issue, we propose a Mutual Information Guided Invertible Image Hiding Network (MIGIIHNet), which leverages mutual information estimation between the lost information and the stego image in the forward process to guide the backward mapping for reconstruction. Specifically, we propose a lightweight INN with a channel attention feature aggregation module (CAFAM), integrating a channel attention mechanism to optimize the multi-scale aggregation of both low-level and high-level features in a single forward pass. Also, an association learning module (ALM) is designed to model the mutual information between the stego image and the lost information during the forward hiding process. Then, the mutual information is utilized to reconstruct the secret image with high accuracy. Extensive experimental results show that MIGIIHNet outperforms existing state-of-the-art methods in terms of invisibility, security, and recovery accuracy, while maintaining low computational complexity.}, archive = {J_EAAI}, author = {Kehan Zhang and Fen Xiao and Jingwen Cai and Xieping Gao}, doi = {10.1016/j.engappai.2025.112343}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112343}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Mutual information guided invertible image hiding network}, volume = {162}, year = {2025}, }
• (2025). Long-tailed detection and classification of wafer defects from scanning electron microscope images robust to diverse image backgrounds and defect scales. EAAI, 162, 112342. (www)

  @article{ , abstract = {In semiconductor engineering, high yield of wafers relies on accurate detection and classification of wafer defects. The dataset for detecting wafer defects presents three primary challenges: (i) different background types, (ii) variable image or defect scales, and (iii) imbalanced data with a long-tailed distribution of defect types. These challenges create significant limitations for traditional classification techniques. To address these issues, we propose a stratified framework called Wafer Detection and Classification (WaferDC), designed specifically for detecting and classifying wafer defects from scanning electron microscope (SEM) images. Our framework achieves high defect detection performance on SEM wafer images by utilizing a multi-cluster memory bank, which effectively handles the challenges of (i) variable background types and (ii) differing image or defect scales. Building on this robust detection, we propose Segmentation and Mix (SegMix), a novel defect augmentation technique based on anomaly heatmaps, which enhances the reliability of defect detection and classification in a (iii) long-tailed imbalanced environment. Finally, we pass defect-classified images through a parameter-efficient fine-tuning (PEFT)-based classifier (Shiet al., 2023) utilizing a vision transformer (ViT) architecture, further improving overall defect detection and classification performance. We rigorously tested WaferDC on a proprietary SEM wafer dataset and the public Describable Textures Dataset-Synthetic (DTD-Synthetic) and Magnetic Tile Defect (MTD) datasets. The results confirm the effectiveness of our method in improving defect detection and classification in wafer manufacturing. Our code is available at https://github.com/SpatialAILab/WaferDC .}, archive = {J_EAAI}, author = {Taekyeong Park and Yongho Son and Sanghyuk Moon and Seungju Han and Je Hyeong Hong}, doi = {10.1016/j.engappai.2025.112342}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112342}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Long-tailed detection and classification of wafer defects from scanning electron microscope images robust to diverse image backgrounds and defect scales}, volume = {162}, year = {2025}, }
• (2025). Fast trajectory planner with a reinforcement learning-based controller for robotic manipulators. EAAI, 162, 112341. (www)

  @article{ , abstract = {Generating obstacle-free trajectories for robotic manipulators in unstructured and cluttered environments remains a significant challenge. Existing motion planning methods often require additional computational effort to generate the final trajectory by solving kinematic or dynamic equations. This paper highlights the strong potential of model-free reinforcement learning methods over model-based approaches for obstacle-free trajectory planning in joint space. We propose a fast trajectory planning system for manipulators that combines vision-based path planning in task space with reinforcement learning-based obstacle avoidance in joint space. We divide the framework into two key components. The first introduces an innovative vision-based trajectory planner in task space, leveraging the large-scale fast segment anything (FSA) model in conjunction with basis spline (B-spline)-optimized kinodynamic path searching. The second component enhances the proximal policy optimization (PPO) algorithm by integrating action ensembles (AE) and policy feedback (PF), which greatly improve precision and stability in goal-reaching and obstacle avoidance within joint space. These proximal policy optimization (PPO) enhancements increase the algorithm’s adaptability across diverse robotic tasks, ensuring consistent execution of commands from the first component by the manipulator, while also enhancing both obstacle avoidance efficiency and reaching accuracy. The experimental results demonstrated the effectiveness of proximal policy optimization (PPO) enhancements, as well as simulation-to-simulation (Sim-to-Sim) and simulation-to-reality (Sim-to-Real) transfer, in improving model robustness and planner efficiency in complex scenarios. These enhancements allowed the robot to perform obstacle avoidance and real-time trajectory planning in obstructed environments. https://sites.google.com/view/ftp4rm/home}, archive = {J_EAAI}, author = {Yongliang Wang and Hamidreza Kasaei}, doi = {10.1016/j.engappai.2025.112341}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112341}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fast trajectory planner with a reinforcement learning-based controller for robotic manipulators}, volume = {162}, year = {2025}, }
• (2025). Bi-level location-routing problem with time windows for mixed-load recycling of heterogeneous batteries: A transformer-based improved deep reinforcement learning algorithm. EAAI, 162, 112340. (www)

  @article{ , abstract = {Given the urgency, timeliness and uncertainty of power batteries recycling, we extend a novel network model for retired battery recycling systems, bi-level capacitated location-routing problem with time windows for heterogeneous battery mixed-load (CLRPTW-HBM). Firstly, the expected interval and linear weighting methods are used to transform and process the lower-level multi-objective function that contains fuzzy variables. Secondly, a Transformer-based improved deep reinforcement learning algorithm (Transformer-IDRL) is proposed. (1) The bi-level CLRPTW-HBM is modeled as Markov decision process, and a policy network model with dual-layer encoder-decoder structure is designed based on Transformer architecture. (2) Randomly generate instance data, and the asynchronous advantage Actor-Critic with adaptive dynamic parameter tuning strategy is employed for training. (3) The action sampling strategy based on roulette reverse selection mechanism, and local search strategy incorporating problem characteristics are introduced to improve solution quality. Finally, extensive experiments are conducted on benchmark datasets and actual cases, and the results demonstrate superior performance of Transformer-IDRL, with an average Gap of 0.22 % and 0.19 %, a 5.30 % reduction in recycling cost, and a 6.06 % reduction in battery exposure risk. These satisfactory results highlight the feasibility and efficiency of the proposed model and method. Additionally, sensitivity analysis of model parameters shows that under different decision-maker preferences and vehicle loading capacity, the sensitivity range of path cost is [5.77 %, 39.85 %] and [16.04 %, 35.54 %], while that of exposure risk is [1.40 %, 3.39 %] and [1.35 %, 5.54 %], indicating that parameter variations significant influence the layout of recycling network and overall path cost. Therefore, decision-makers should flexibly adjust key parameters to balance economic benefits and sustainable development in battery recycling.}, archive = {J_EAAI}, author = {Mengna Zhao and Shiping Chen}, doi = {10.1016/j.engappai.2025.112340}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112340}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Bi-level location-routing problem with time windows for mixed-load recycling of heterogeneous batteries: A transformer-based improved deep reinforcement learning algorithm}, volume = {162}, year = {2025}, }
• (2025). Instantaneous power prediction for industrial robots using tree-based machine learning methods. EAAI, 162, 112339. (www)

  @article{ , abstract = {This article presents a tree-based machine learning methodology for instantaneous power prediction designed, tested and validated using data from an articulated industrial robot. The proposed methodology for instantaneous power prediction materializes through a generic system architecture with functionalities consisting of data acquisition, time alignment of data samples, storage, model learning, instantaneous power prediction and integration in time to evaluate energy consumption at robot operation level. This methodology is designed to evaluate offsite energy consumption of robotized workstations for different layouts characterized by relative position of the robot with respect to the serviced and fly-by points. This is important both for offline virtual commissioning of robotized workstations (determine layout) and for online operation for maintenance purposes (determine energy spikes different from normal model). The analyzed operation is the linear motion of the robot Tool Control Point in Cartesian space, characterized by the complexity of the kinematic model: each joint operates in coordinated motion, adjusting its velocity and acceleration continuously to ensure a straight path with constant speed. A custom Internet of things (IoT) device enables synchronized energy and motion data logging for robots, ensuring consistent values for sampled trajectories. Justification for the usage of tree-based methods and experimental results are provided.}, archive = {J_EAAI}, author = {Ionuţ Lenţoiu and Silviu Răileanu and Theodor Borangiu and Mihnea Constantinescu and Octavian Morariu}, doi = {10.1016/j.engappai.2025.112339}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112339}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Instantaneous power prediction for industrial robots using tree-based machine learning methods}, volume = {162}, year = {2025}, }
• (2025). Toward sustainable wastewater treatment: Transformer ensembles and multitask learning for energy consumption and quality management. EAAI, 162, 112338. (www)

  @article{ , abstract = {Wastewater treatment plants (WWTPs) are among the most energy-intensive components of urban infrastructure and bear strict regulatory responsibilities for wastewater quality. These dual challenges, minimizing energy consumption and maintaining environmental compliance, are deeply interrelated and must be managed simultaneously to achieve sustainable plant operation. This study proposes a framework that comprises two customized components. The first component employs a voting ensemble model based on transformer architecture to predict energy consumption. It processes heterogeneous feature domains — including hydraulic, wastewater, and climatic variables — through parallel attention-driven streams. The outputs from these streams are then aggregated using a weighted voting mechanism to produce the final prediction. Second, a multitask Bidirectional Gated Recurrent Unit (Bi-GRU) forecasts wastewater quality indicators concurrently (ammonia, Biochemical Oxygen Demand (BOD), and Chemical Oxygen Demand (COD)), capturing shared temporal dependencies and reducing model complexity. A hybrid preprocessing strategy is applied, incorporating domain-aware outlier detection (z-score and Interquartile Range (IQR)), K-Nearest Neighbors (KNN) Imputation, and feature selection using Extreme Gradient Boosting (XGBoost). Experimental results showed that. The voting ensemble model achieved the best results for energy consumption prediction with 31.61 of Root Mean Squared Error (RMSE). The multitask Bi-GRU achieved the best results for wastewater quality indicators with RMSE at 6.1689, 48.0323, and 88.2214 for ammonia, BOD, and COD, respectively. This work is among the first to integrate transformer ensembles and multitask learning in a unified WWTP forecasting system. Simultaneously addressing energy efficiency and water quality assurance, this offers a practical, scalable, and intelligent decision-support tool for sustainable wastewater management.}, archive = {J_EAAI}, author = {Hager Saleh and Sherif Mostafa and Shaker El-Sappagh and Abdulaziz AlMohimeed and Michael McCann and Saeed Hamood Alsamhi and Niall O’Brolchain and John G. Breslin and Marwa E. Saleh}, doi = {10.1016/j.engappai.2025.112338}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112338}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Toward sustainable wastewater treatment: Transformer ensembles and multitask learning for energy consumption and quality management}, volume = {162}, year = {2025}, }
• (2025). Complete information extraction for monocular depth estimation using a dual framework. EAAI, 162, 112337. (www)

  @article{ , abstract = {This paper aims to address the problem of efficient extraction of complete multi-scale information for supervised monocular depth estimation. Most of the existing depth estimation methods are based on Convolutional Neural Network (CNN). By gradually exploring the contextual and semantic features, they have achieved good results in scene depth estimation. However, with the expansion of the receptive field, global information limited by the local induction bias is gradually suppressed, resulting in the performance cannot be further improved. Recently, Transformer-based methods have been widely used to model the global correlation between features. Nevertheless, since the Transformer networks are not spatially aware enough, they usually lose local details and have no clear mechanism for reusing features when processing images. The Transformer networks perform self-attention mechanism at each location and cannot directly obtain information from other locations for features. Therefore, we propose a novel dual framework called as Transformer-CNN, which includes the Transformer-branch and the CNN-branch for monocular depth estimation. Specifically, the Transformer-branch is able to model the global contextual information and the CNN-branch can capture local spatial relationships in images. However, simply fusing these two independent branches may result in insufficient feature aggregation. To this end, we design a Parallel Feature Interaction Module (PFIM), which contains a Self-Attention Module (SAM) and a Cross-Attention Module (CAM), so as to highlight features from the Transformer-branch and the CNN-branch respectively and extract complementary information between the two branches. Meanwhile, in order to make full use of the low-level features with low quality in the scene, we propose a Low-level Information Acquisition Module (LIAM) to capture texture-related information and preserve texture details in the CNN-branch. Finally, to address the lack of multi-scale contextual information in Vision Transformer (ViT), we introduce a Wide Area Multi-scale Decoder (WAMD), which incorporates the multi-scale feature representations into the decoder part via a Wide Area Attention (WAA). Extensive experiments on benchmark datasets collected in the outdoor and indoor environments demonstrate the competitive results of the proposed method, compared with the state-of-the-art monocular depth estimation methods.}, archive = {J_EAAI}, author = {Bin Li and Dazheng Zhou and Xianjie Gao and Mingliang Zhang}, doi = {10.1016/j.engappai.2025.112337}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112337}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Complete information extraction for monocular depth estimation using a dual framework}, volume = {162}, year = {2025}, }
• (2025). Towards non-visual bowel cancer diagnosis: A certainty-aware data-driven method of lesion characterisation using a vibrating capsule. EAAI, 162, 112336. (www)

  @article{ , abstract = {Recent advances in miniaturised, dynamically actuated robots have opened new pathways for non-visual, in-situ disease diagnosis. This study explores a novel method for early bowel cancer detection using a self-propelled robotic capsule that navigates the bowel and detects lesions based on variations in tissue stiffness. The approach capitalises on the sensitivity of the capsule’s dynamic responses to surrounding tissue properties. A dual-phase machine learning framework is proposed. The first phase uses regression models including multilayer perceptron (MLP), support vector regression (SVR), and Gaussian process regression (GPR) to predict tissue stiffness from displacement signal features. The second phase uses a Gaussian mixture model (GMM) to cluster the predicted stiffness values into different categories. Unlike our previous work, this study emphasises the robustness of the models under varying data conditions using both accuracy and reliability-oriented metrics. Based on our studies, MLP provided the most reliable regression results for simulated data and downstream clustering, though GPR performed better on experimental datasets. SVR consistently underperformed, especially on experimental data. The GMM achieved over 89% clustering accuracy across both simulated and experimental datasets, with improved results when predictions from more accurate regression models are used as the inputs. This work demonstrates a promising step toward dynamic, in-situ lesion characterisation and highlights the potential for integrating lesion biomechanics into future endoscopic diagnosis.}, archive = {J_EAAI}, author = {Kenneth Omokhagbo Afebu and Yang Liu and Evangelos Papatheou and Shyam Prasad}, doi = {10.1016/j.engappai.2025.112336}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112336}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Towards non-visual bowel cancer diagnosis: A certainty-aware data-driven method of lesion characterisation using a vibrating capsule}, volume = {162}, year = {2025}, }
• (2025). Beyond the horizon: A comprehensive analysis of artificial intelligence-based weather forecasting models. EAAI, 162, 112335. (www)

  @article{ , abstract = {The field of Artificial Intelligence (AI)-based weather forecasting is growing rapidly, with continuous progress in model development, techniques, and performance improvements. This paper provides a comprehensive overview of AI-based weather forecasting models, focusing on their current status, challenges, and directions for further development. A review of more than 40 models, primarily proposed after 2015, underscores the importance of critically examining various aspects of AI-based forecasting. Unlike previous reviews that targeted only a limited number of models or features, this study addresses a complete set of aspects and analyzes existing challenges from multiple perspectives. These aspects include the Machine Learning (ML) and Deep Learning (DL) methods used, datasets, predictand parameters, overfitting, and capability for forecasting extreme weather, lead time, spatiotemporal scale, performance criteria, overfitting, data assimilation, data-driven models, and the analysis of state-of-the-art (SOTA) models such as FengWu, ClimaX, Pangu-Weather, FourCastNet, GraphCast, GenCast, and Artificial Intelligence Forecasting System (AIFS) from various viewpoints. The review also discusses current challenges, including limited historical data and data quality, small-scale weather forecasting, model explainability, uncertainty, extreme weather prediction, physical constraints, temporal adaptation, and generalization, and outlines potential future directions.}, archive = {J_EAAI}, author = {Saeid Haji-Aghajany and Witold Rohm and Piotr Lipinski and Maciej Kryza}, doi = {10.1016/j.engappai.2025.112335}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112335}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Beyond the horizon: A comprehensive analysis of artificial intelligence-based weather forecasting models}, volume = {162}, year = {2025}, }
• (2025). Neural model of the adaptive tuned particle impact damper. EAAI, 162, 112334. (www)

  @article{ , abstract = {The paper presents a novel approach for the modeling of the Adaptive Tuned Particle Impact Damper (ATPID) using Multilayer Perceptron (MLP). The main motivation was the recognition that such an approach can support the development of novel neural modeling and the optimal determination of damper parameters in terms of mechanical vibration attenuation. The training data were obtained using a theoretical model validated experimentally. The optimally selected MLP was compared with other regression models using 10 different metrics. A hyperparameter tuning of the determined neural network architecture was conducted based on the input parameters such as excitation amplitude, grain mass, and ATPID damper height. The analyses show that the proposed neural network could quickly and accurately estimate the system’s vibration amplitude and efficiently predict the optimal damper height. The ability to effectively determine the correct optimal height is crucial for ATPID damper control. The high efficiency in predicting the system’s vibration amplitude allows for the replacement of the theoretical model with applied time-consuming contact forces. The MLP accurately estimated vibration amplitudes with 1%–10% error for interpolated data and up to 15% for extrapolated cases. The issue raised is particularly important from the perspective of real-time damper control. It was found that computing a single case using the artificial neural network is more than ten times faster compared to the theoretical model. Therefore, the proposed ATPID damper model based on a neural network forms the basis for further considerations and scientific research to finally propose a control algorithm in the future.}, archive = {J_EAAI}, author = {Mateusz Żurawski and Karolina Grabska and Robert Zalewski and Adam Kulawik}, doi = {10.1016/j.engappai.2025.112334}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112334}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Neural model of the adaptive tuned particle impact damper}, volume = {162}, year = {2025}, }
• (2025). An adaptive hybrid machine reading comprehension framework for multimodal emotion-cause pair extraction in conversations. EAAI, 162, 112333. (www)

  @article{ , abstract = {Multimodal emotion-cause pair extraction in conversations (MECPEC) has gradually evolved into an emerging task aimed at discovering deeper causal relationships between emotions and their corresponding causes in conversational contexts. It has widespread application in fields such as human–computer interaction, social media analysis, customer feedback management, and empathetic companion, among others. However, the challenges posed by the oversimplified multimodal feature fusion mechanism and the failure to account for the relative positional relationship between emotions and causes still hinder the performance of MECPEC. In this study, we propose an adaptive hybrid machine reading comprehension (AHMRC) framework to extract potential emotion-cause pairs inherent in conversations. The MECPEC task is first transformed into a two-round hybrid machine reading comprehension task that sequentially enforces the global emotion query and the local cause query with the goal of exploring the relative position constraint specific to conversations. Subsequently, an adaptive multimodal attention module is designed by incorporating features extracted from text, video, and audio modalities, and adaptively fusing them according to their contributions. Extensive experiments were carried out on the benchmark datasets to demonstrate the effectiveness of the proposed AHMRC framework in comparison to other state-of-the-art methods in the literature.}, archive = {J_EAAI}, author = {Guorui Li and Xufeng Duan and Cong Wang and Sancheng Peng}, doi = {10.1016/j.engappai.2025.112333}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112333}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An adaptive hybrid machine reading comprehension framework for multimodal emotion-cause pair extraction in conversations}, volume = {162}, year = {2025}, }
• (2025). Analytical and machine learning approaches for three-dimensional orthotropic functionally graded cylindrical structure in steady heat conduction. EAAI, 162, 112332. (www)

  @article{ , abstract = {Theoretical and computational modelling of heat conduction in a functionally graded cylinder has been rigorously investigated in prior studies, owing to its critical significance in high-stakes engineering applications such as nuclear reactor design, aerospace structural systems, and pressure vessel technology. Despite this extensive body of work, the majority of these studies have mostly focused on simplified two-dimensional models, frequently presuming idealized thermal parameters such as isotropic thermal conductivity, constant convection coefficients, and spatially uniform ambient temperatures. These simplifications overlook the crucial role of spatially varying thermal characteristics and three-dimensional (3D) temperature distributions, which are essential for accurately simulating the complex heat conduction behaviors found in real-world engineering. Unlike prior research, this study presents an analytical solution for heat conduction under non-homogeneous generalized Robin boundary conditions, capturing 3D thermal conductivity inhomogeneities along three orthogonal directions using Sturm-Liouville theory and finite integral transforms. However, while such analytical methods are highly accurate for simpler geometries, they often encounter significant challenges when extended to scenarios involving complex material gradients and irregular domains. A gradient-enhanced physics-informed neural network (g-PINN) framework is proposed to tackle these challenges, utilizing neural networks that incorporate physical laws and gradient information to enhance its applicability to complex configurations. This combined approach presents a novel framework that integrates classical theory with machine learning, facilitating precise modelling of thermal phenomena in functionally graded cylinders. The findings indicate that both the analytical and machine learning approaches (g-PINN) align closely with presented solutions, precisely capturing the energy equation and more complex boundary conditions.}, archive = {J_EAAI}, author = {Palash Das and Md Ashraful Islam and Dipayan Mondal}, doi = {10.1016/j.engappai.2025.112332}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112332}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Analytical and machine learning approaches for three-dimensional orthotropic functionally graded cylindrical structure in steady heat conduction}, volume = {162}, year = {2025}, }
• (2025). Robust spurious drug–drug interaction detection via chi-square-guided bidirectional attention mechanism. EAAI, 162, 112331. (www)

  @article{ , abstract = {Spurious drug–drug interactions arising from experimental biases can compromise treatment safety and hinder effective clinical decision-making for patients. However, current molecular representation learning methods still face significant challenges in addressing this problem. First, most graph neural network-based approaches consider only single-direction interactive semantic extraction between drug molecules, failing to capture more granular relationships between drugs fully. Moreover, during training, they are vulnerable to noise from negative sampling and cannot sufficiently leverage the complete drug–drug interaction annotations, resulting in reduced robustness and accuracy in downstream tasks. To this end, we propose a robust spurious DDI detection framework that employs chi-square-guided bidirectional attention to capture fine-grained and bidirectional interaction patterns. First, considering their mutual information flow, a two-way cross-attention mechanism is introduced for a more granular extraction of cross-drug molecular interaction semantic representations through bidirectionally perceiving interactive features between drugs. Second, on the basis of robust minimum covariance determinant theory, we propose a chi-square distribution-based spurious detection method to approximate the correctly annotated drug–drug interactive feature space to a chi-square distribution for a complete feature representation. Extensive experiments on benchmark datasets further validate our method’s effectiveness over state-of-the-art methods, particularly in noisy interference scenarios. Our code is available at https://github.com/AlexCostra/cd .}, archive = {J_EAAI}, author = {Wei-Yu Shi and Yi-Jia Zhang and Jin-Zhong Ning}, doi = {10.1016/j.engappai.2025.112331}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112331}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Robust spurious drug–drug interaction detection via chi-square-guided bidirectional attention mechanism}, volume = {162}, year = {2025}, }
• (2025). A domain adaptation method to defend chinese textual adversarial attacks via prompt-tuning. EAAI, 162, 112330. (www)

  @article{ , abstract = {The textual adversarial attack aims to fool existing models into making erroneous predictions by adding strategic perturbations to normal data without affecting the user’s understanding. Recently, methods based on Pre-trained Language Models (PLMs) and Large Language Models (LLMs) have shown promising performance in various Natural Language Processing (NLP) downstream tasks. However, due to significant deviations between the original and perturbed texts, these methods struggle to achieve satisfactory results in defending against textual adversarial attacks, especially in Chinese, which has unique syntactic structures. To address this issue, we propose a domain adaptation method for defending against Chinese textual adversarial attacks through a prompt-tuning model, which effectively mitigates the discrepancy between different domains. Specifically, the original and perturbed texts are treated as the source and target domains, respectively, with the textual adversarial defense task framed as a cross-domain classification problem. The soft prompt-tuning model trained in the source domain is iteratively adapted to uncover the true label information in the target domain. The graph attention network is incorporated to integrate Chinese syntactic structure information with semantic features. Through a voting mechanism on predicted labels generated by the iterative model, soft prompt-tuning is further optimized for cross-domain classification tasks. Extensive experimental results demonstrate the superior effectiveness of our method in Chinese textual adversarial defense tasks compared to baseline methods, including the state-of-the-art fine-tuning approaches for PLMs and LLMs.}, archive = {J_EAAI}, author = {Yi Zhu and Zhenglong Li and Yun Li and Yunhao Yuan and Jipeng Qiang}, doi = {10.1016/j.engappai.2025.112330}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112330}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A domain adaptation method to defend chinese textual adversarial attacks via prompt-tuning}, volume = {162}, year = {2025}, }
• (2025). A novel high-accuracy graph neural network-based rumor detection method. EAAI, 162, 112329. (www)

  @article{ , abstract = {Rumors spreading on social media platforms result in potential damages. A precise rumor detection mechanism can help form a healthy public opinion environment. In recent years, deep learning-based rumor detection methods, especially graph model-based ones, have risen and reached promising performance. However, there are several defects in existing methods, which limit models from efficiently utilizing the propagation structure. In this paper, we propose a novel rumor detection model, which has high accuracy and reaches state-of-the-art performance. First, we design a powerful comprehensive rumor feature extractor that explicitly overcomes the restriction of previous Graph Neural Networks-based models. Then, by introducing Kernel Subtree features, our model acquires the capability to learn crucial local features from important nodes. Comparative experiments performed on two real-world social media platforms demonstrate that our work reaches state-of-the-art performance, which outperforms the best baseline with 1.6% and 1.9% in accuracy respectively.}, archive = {J_EAAI}, author = {Xi Xiao and Zeming Wu and Chengzong Cai and Tian Bian and Guangwu Hu and Qing Li and Cheng Huang}, doi = {10.1016/j.engappai.2025.112329}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112329}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel high-accuracy graph neural network-based rumor detection method}, volume = {162}, year = {2025}, }
• (2025). Committee of multi-scale nonlinear learning frameworks for accurate stock price forecasting. EAAI, 162, 112325. (www)

  @article{ , abstract = {Dependable stock price predictions are vital for optimizing economic policies and investment strategies in both national and corporate settings. However, the intrinsic volatility and intricacy of stock prices pose considerable challenges. Thus, this paper introduces a novel Committee of Multi-scale Nonlinear Learning Frameworks (CoML) that employs a three-stage model: decomposition, reconstruction, and prediction. First, a complete ensemble empirical mode decomposition with adaptive noise is adopted to decompose the original stock prices into multiple intrinsic mode functions. Secondly, a fine-to-coarse algorithm is applied to reconstruct the intrinsic mode functions, so as to effectively extract short-term fluctuations and long-term trends. Finally, an ensemble of nonlinear models including bidirectional long short-term memory (BiLSTM), support vector regression (SVR) and multi-layer perceptron (MLP) is used to learn and forecast features extracted to obtain high performance. Experimental results indicate that the model performs exceptionally well in both emerging and developed markets highlighting the innovative capabilities of CoML in highly complex and volatile financial markets. The proposed model is further validated using Model Confidence Set and the results indicate that the model is statistically significant.}, archive = {J_EAAI}, author = {Qian He and Yanhui Liang and Yu Lin and Dazhi Pan and Yuying Yue}, doi = {10.1016/j.engappai.2025.112325}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112325}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Committee of multi-scale nonlinear learning frameworks for accurate stock price forecasting}, volume = {162}, year = {2025}, }
• (2025). Learning robust brain tumor segmentation under label corruption and data scarcity. EAAI, 162, 112322. (www)

  @article{ , abstract = {Deep learning models for medical image segmentation often struggle with performance issues when datasets are affected by label noise or annotation errors, commonly introduced during manual process or lack of proficiency. These noisy annotations disrupt the loss function, leading to "partially incorrect" gradients that impair the model's learning and overall performance. Additionally, the limited availability of scanned data for training often makes it challenging to develop a robust model. A common approach to address this issue is to leverage similar large annotated datasets. However, differences in dataset distributions can also lead to inconsistencies, introducing erroneous gradients during training and further impacting model performance. To address these challenges, we propose MGR-DAS (Meta-Gradient Reweighting via Direction-Aware Similarity), a novel meta-learning-based approach that can automatically evaluate the reliability of training samples during training using a small, clean subset easily curated from the noisy dataset. Our method quantifies reliability by measuring the cosine similarity between the gradients of noisy training samples and those of the clean subset. Samples with higher gradient alignment are assigned greater weights during training, effectively reducing the impact of noisy labels and improving model robustness. We evaluate our method using three standard metrics for medical image segmentation: the Dice Similarity Coefficient (DSC), the 95th percentile Hausdorff Distance (HD95), and Intersection over Union (IoU). The proposed MGR-DAS achieved an overall 2.4 % improvement in the DSC on the brain tumor segmentation (BraTS, 2021) dataset. Remarkably, even with only 10 clean annotations used in the reweighting algorithm, our method yielded a 28.7 % gain in DSC. In real-world, data-scarce scenarios, our proposed MGR-DAS also improved the overall DSC score by 2.6 % on BraTS pediatric (BraTS-PEDs) and by 1.0 % on BraTS-Africa, demonstrating strong generalizability and robustness. Experimental results confirm that the proposed method reliably identifies noisy data, prioritizes clean data through adaptive weighting, and outperforms existing fine-tuning, curriculum learning techniques, and other meta-learning frameworks commonly employed in classification tasks.}, archive = {J_EAAI}, author = {Abdulkhalek Al-Fakih and Abbas Mohamed Rezk and Abdullah Shazly and Kanghyun Ryu and Mohammed A. Al-masni}, doi = {10.1016/j.engappai.2025.112322}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112322}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Learning robust brain tumor segmentation under label corruption and data scarcity}, volume = {162}, year = {2025}, }
• (2025). Zero-shot learning augmented slow feature analysis for semantic-aware industrial process fault detection. EAAI, 162, 112321. (www)

  @article{ , abstract = {Slow Feature Analysis (SFA) has shown considerable success in the field of industrial process fault detection. Nonetheless, due to its unsupervised nature, SFA relies solely on the normal training data and overlooks the incorporation of prior process knowledge, which consequently diminishes its efficacy in early fault detection. To mitigate this limitation, this paper introduces the concept of Zero-Shot Learning (ZSL) and proposes an improved SFA approach, referred to as ZSL-SFA. This novel method leverages fault semantic representations as auxiliary knowledge to enhance fault detection sensitivity in industrial process monitoring. The ZSL-SFA framework implements a dual-model collaborative monitoring system: (1) a primary SFA model is developed using normal operational data to capture the dynamic characteristics of the process; and (2) a semantic encoding mechanism, grounded in expert knowledge, is devised to build the auxiliary model, where a probabilistic attribute learner adaptively extracts semantic information from fault attribute descriptions, facilitating effective fault knowledge transfer through similarity analysis. The monitoring outcomes from both the primary and auxiliary models are integrated using a Bayesian fusion strategy, culminating in a comprehensive ZSL-SFA monitoring system. The main advantage of this method is its ability to fully exploit prior process knowledge to enhance the basic SFA model without the need for additional labeled fault samples. Experimental validations on the Tennessee-Eastman process simulation platform are performed to indicate that the proposed ZSL-SFA method surpasses the basic SFA method in terms of fault detection performance.}, archive = {J_EAAI}, author = {Wenjie Yang and Xiaogang Deng and Lumeng Huang and Yuping Cao}, doi = {10.1016/j.engappai.2025.112321}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112321}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Zero-shot learning augmented slow feature analysis for semantic-aware industrial process fault detection}, volume = {162}, year = {2025}, }
• (2025). Multi-scale embedding with guided attention for medical image analysis. EAAI, 162, 112319. (www)

  @article{ , abstract = {Accurate classification of medical images is pivotal for enhancing diagnostic accuracy and optimizing treatment strategies. Traditional methods encounter challenges in delineating clear inter-class boundaries within high-dimensional feature spaces affected by class overlap. This study introduces a Multi-Scale Embedding with Guided Attention (MSEGA) framework based on deep learning autoencoders that integrates innovative guided attention learning mechanisms without explicit target mask supervision for tumor detection and classification. The framework incorporates Multi-scale Feature Extraction Blocks and Depthwise Separable Convolution Blocks to comprehensively capture image features. Through Channel Attention and Spatial Attention mechanisms, the MSEGA method prioritizes critical tumor regions across scales. We also propose a novel interpretable embedding learning loss function to optimize image embeddings, highlighting crucial regions and refining category distinctions. Empirical evaluations on two brain tumor Magnetic Resonance Imaging (MRI) datasets demonstrate our approach surpasses conventional methods including Convolutional Neural Networks and Vision Transformers in classification accuracy and generalizability. These results underscore the framework’s potential as a versatile artificial intelligence-powered tool in medical image analysis.}, archive = {J_EAAI}, author = {Zeyan Li and Yifei Peng and Yizun Lin}, doi = {10.1016/j.engappai.2025.112319}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112319}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-scale embedding with guided attention for medical image analysis}, volume = {162}, year = {2025}, }
• (2025). Multimodality based deep learning method for cancer-related T-cell receptor sequence prediction. EAAI, 162, 112318. (www)

  @article{ , abstract = {T-cell receptor sequences (TCR-seq) are closely related to cancers, and in particular, cancer-related TCR-seq are crucial in cancer diagnosis and treatment. Current prediction methods for cancer-related TCR-seq often focus solely on the sequence structure, neglecting its spatial structure. Therefore, we propose a multimodal deep learning method based on parallel and residual structures (MDPR) for the detection of cancer-related TCR-seq. MDPR can effectively integrate the spatial and sequence structure of TCR-seq for accurately identifying cancer-related sequences. First, we introduce a TCR-seq encoding method based on atomic three-dimensional spatial coordinates, allowing for more effective extraction of the spatial structural features of TCR-seq. Second, we use high-dimensional word vectors instead of the amino acid feature vectors traditionally used by other researchers. Third, we pretrain the spatial feature extraction module and then conduct joint training with the sequence feature extraction module. This approach allows the model to better consider the relationship between the two modalities, thereby improving prediction accuracy. Finally, MDPR achieved an area under the curve (AUC) of 0.971 after ten rounds of three-fold cross-validation on the dataset. The AUC of MDPR is 5% higher than that of the previous best method. In short, we propose an artificial intelligence method called MDPR, and apply it to the biomedical field. MDPR can be obtained from https://github.com/biomg/MDPR .}, archive = {J_EAAI}, author = {Junjiang Liu and Shusen Zhou and Mujun Zang and Chanjuan Liu and Tong Liu and Qingjun Wang}, doi = {10.1016/j.engappai.2025.112318}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112318}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multimodality based deep learning method for cancer-related T-cell receptor sequence prediction}, volume = {162}, year = {2025}, }
• (2025). A combined perspective self-supervised contrastive learning framework for human activity recognition integrating instance prediction and clustering. EAAI, 162, 112317. (www)

  @article{ , abstract = {Human activity recognition (HAR) is a significant application for wearable devices, which primarily identifies current human activities by analyzing sequential sensor data. The real-time data recording of wearable devices enables the collection of vast amount of unlabeled data. Utilizing this data for self-supervised contrastive pre-training of HAR models presents a feasible solution to the decline in recognition performance due to limited labeled data. However, traditional contrastive learning frameworks are primarily designed based on positive and negative sample pairs in the image domain. The relatively simple sequence data of HAR is prone to generating incorrect negative pairs, thus pre-training HAR models solely in this manner is unsatisfactory. Given the phenomena described above, this paper proposes an Instance Prediction and Clustering Self-supervised Contrastive Learning Framework (IPCSC) for HAR, considering the characteristics of human activity data. IPCSC circumvents negative sample pairs, instead extracting contrastive information at the instance perspective by prediction tasks among various augmented views of samples and integrating clustering concepts for contrastive learning from a holistic perspective. The primary objective is to enable the model to discern critical information within human activity data and distinguishable features between different activities, thereby improving the model’s pre-training efficacy and enhancing its downstream activity recognition performance. Numerous experimental analyses demonstrate that IPCSC outperforms other self-supervised methods, achieving an average F1-Score performance improvement of 5.65%, 4.11%, and 7.99% over supervised baselines on the UCI-HAR, MobiAct, and MotionSense datasets, respectively, with only 1% of the labeled data.}, archive = {J_EAAI}, author = {Zhixuan Yang and Kewen Li and Zongchao Huang and Zhifeng Xu and Xinyuan Zhu and Yuan Xiao}, doi = {10.1016/j.engappai.2025.112317}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112317}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A combined perspective self-supervised contrastive learning framework for human activity recognition integrating instance prediction and clustering}, volume = {162}, year = {2025}, }
• (2025). A self-supervised method for learning path-augmented knowledge graph embedding. EAAI, 162, 112315. (www)

  @article{ , abstract = {Knowledge graphs (KGs) consist of factual triples that describe relations between entities in the real world. Knowledge graph embedding (KGE) aims to map entities and relations into constantly low-dimensional vectors, which is important for lots of downstream tasks (e.g., KG completion and information retrieval). Current KGE methods primarily rely on explicit structural patterns, neglecting latent contextual semantics behind those structures and resulting in sub-optimal performance. While some methods incorporate additional data (e.g., textual descriptions), such dependencies limit applicability due to additional data requirements. Furthermore, most KGE models suffer from limited supervision with sparse labeled triples, restricting their capacity to learn comprehensive semantic features. Inspired by the simple but effective self-supervised language model word2vec, one interesting question is: Can KGE be performed as a simple self-supervised language model ? To achieve this, we innovatively propose a self-supervised KGE framework that learns entity and relation embeddings by adapting word2vec’s skip-gram objective to path sequences extracted from KGs. Our framework employs separate embedding spaces for entities and relations with an entity-relation mapping mechanism for effective interaction between the two embedding spaces. Further, to enhance the training efficiency, we introduce a markov chain-based negative sampling strategy, which generates semantically meaningful negative samples by preserving the structural contexts along KG paths. Our framework, which is the first attempt to follow the context-based self-supervised idea of language models to conduct KGE tasks, addresses the constraints of label-dependent supervised KGE techniques and obviates the requirement for external information, while simultaneously enabling effective extraction of the implicit contextual semantics inherent in triple structures. Experiments on two widely-used KGE datasets show state-of-the-art performance, demonstrating our framework’s ability to learn semantically rich representations solely from graph structure.}, archive = {J_EAAI}, author = {Tong Shen and Fu Zhang and Jingwei Cheng}, doi = {10.1016/j.engappai.2025.112315}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112315}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A self-supervised method for learning path-augmented knowledge graph embedding}, volume = {162}, year = {2025}, }
• (2025). A single-cell RNA sequencing data imputation method based on non-negative matrix factorization and multi-kernel similarity network fusion. EAAI, 162, 112313. (www)

  @article{ , abstract = {Artificial intelligence-based single-cell RNA sequencing (scRNA-seq) technology is widely used in cell type identification and disease research, but its data often contain a large number of missing values and zero values due to technical limitations and biological differences. These zero values not only affect downstream analysis, but also make it difficult to distinguish technical zero values from biological zero values. Therefore, this paper proposes a scRNA-seq data interpolation method (sc-MKNMF) based on non-negative matrix factorization and multi-kernel similarity network fusion for the first time. This method improves the accuracy of cell clustering by accurately filling some zero values. First, sc-MKNMF uses gene-cell dual-level analysis to distinguish technical zero values from biological zero values, and then calculates the similarity network of multi-kernel fusion of genes and cells respectively. Then, this method uses non-negative matrix factorization combined with similarity network to construct the objective function, and introduces sparse regularization terms to ensure the similarity between genes and cells and improve stability. In addition, sc-MKNMF is also equipped with an efficient optimization algorithm to promote its convergence by continuously updating the objective function. Finally, the verification and comparative experiments on 12 scRNA-seq datasets show that the sc-MKNMF method outperforms other advanced data interpolation methods. In addition, the extension of sc-MKNMF to the two tasks of cell trajectory inference and differentially expressed gene analysis showed significant improvement and excellent versatility.}, archive = {J_EAAI}, author = {Pei Liu and Cheng Chen and Hao Liu and Jin Gu and Xinya Chen and Ying Su and Zhiyuan Cheng and Xiaoyi Lv and Chen Chen}, doi = {10.1016/j.engappai.2025.112313}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112313}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A single-cell RNA sequencing data imputation method based on non-negative matrix factorization and multi-kernel similarity network fusion}, volume = {162}, year = {2025}, }
• (2025). A novel object detection model for sugar beet cercospora leaf spot in field scenarios based on large kernel decomposition and spatial channel interaction attention. EAAI, 162, 112311. (www)

  @article{ , abstract = {Cercospora leaf spot (CLS) is a widespread disease that seriously threatens beet yield and sugar quality. Timely detection enables farmers to take early control measures and reduce economic losses. Although artificial intelligence (AI)-based methods are replacing manual inspection in agriculture, CLS detection in complex field environments remains highly challenging due to subtle early-stage symptoms and severe occlusions caused by overlapping leaves and weeds. To address these challenges, this paper presents Cercospora Leaf Spot–You Only Look Once (CLS–YOLO), an enhanced detection model built upon You Only Look Once version 11 (YOLOv11), incorporating novel modules specifically designed for accurate CLS detection under challenging field conditions. To improve the detection of weak and early-stage symptoms, we design the Multi-Scale Large Kernel Decomposition (MSLKD) module, which enhances feature extraction for subtle lesions. Furthermore, we develop the Spatial-Channel Interaction Attention (SCIA) module to mitigate detection errors arising from occlusion and fragmented disease patterns by refining multi-scale feature representations. Experimental results demonstrate CLS–YOLO achieves superior performance, reaching an mAP@0.5 of 73.6% ± 0.2% and an mAP@0.5:0.95 of 40.6% ± 0.3% over five independent runs, outperforming twelve mainstream object detection algorithms while maintaining lightweight efficiency. To validate generalization capability across scenarios, crops, and diseases, we conducted comparative experiments on two public crop disease datasets, where our method achieved superior overall performance. In summary, this study provides an effective AI-driven solution for precise crop disease detection, contributing to the practical advancement of intelligent agriculture.}, archive = {J_EAAI}, author = {Hualong Dong and Yi Lu and Yurong Qian and Xuefei Ning and Ting Chen and Ke Tang}, doi = {10.1016/j.engappai.2025.112311}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112311}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel object detection model for sugar beet cercospora leaf spot in field scenarios based on large kernel decomposition and spatial channel interaction attention}, volume = {162}, year = {2025}, }
• (2025). ElecBench: A large language model benchmark in electric power domain. EAAI, 162, 112310. (www)

  @article{ , abstract = {Large language models (LLMs) have made substantial advancements in the field of natural language processing, necessitating the development of new benchmarks to accurately track their progress. In this paper, we introduce ElecBench, the first benchmark specifically designed for the electric power domain. ElecBench comprises 24 datasets spanning different scenarios, covering general electric power knowledge and four specific business applications, with a total of 34,030 data entries. Furthermore, we evaluate the performance of a series of open-source Chinese LLMs on ElecBench. Our experiments demonstrate that ElecBench serves as an effective benchmark for electric power scenarios and highlight that existing LLMs require further optimization to gain domain-specific knowledge and achieve better performance.}, archive = {J_EAAI}, author = {Sai Zhang and Qiaochu Huang and Qiang Zhang and Xiao Liang and Weiwei Liu and Kunlun Gao and Fei Zhou and Congcong Shi}, doi = {10.1016/j.engappai.2025.112310}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112310}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {ElecBench: A large language model benchmark in electric power domain}, volume = {162}, year = {2025}, }
• (2025). Reinforcement learning-based trajectory tracking optimal control for underactuated unmanned surface vehicles under asymmetric input saturation. EAAI, 162, 112307. (www)

  @article{ , abstract = {For underactuated unmanned surface vehicles (USVs) under asymmetric input saturation caused by thrust-limit characteristics, as well as unknown dynamics and ocean environmental disturbances, a trajectory tracking optimal control (TTOC) scheme is proposed using the reinforcement learning (RL) method. Through coordinate transformations and mathematical derivation, an underactuated USV motion model is transformed into the standard affine nonlinear form. To address the asymmetric input saturation of underactuated USVs, a new inverse hyperbolic tangent-type penalty function is designed for control inputs, relaxing the assumption of input saturation limits being symmetric. Based on RL methods and adaptive neural networks (NNs), an actor-critic NN framework is developed, with weight update laws designed for NNs. This framework learns the TTOC law for underactuated USVs through the online interaction of actor and critic NNs while adapting to unknown dynamics and disturbances. In particular, a robustifying term is designed and added to the output of an actor NN to compensate for the adverse effects of a lumped residual term, which enhances the robustness of the TTOC law and thereby achieves asymptotic regulation of trajectory tracking errors. Theoretical analyses and simulation results indicate that the proposed TTOC scheme enables underactuated USVs to asymptotically track the desired trajectory.}, archive = {J_EAAI}, author = {Ziping Wei and Jialu Du}, doi = {10.1016/j.engappai.2025.112307}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112307}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Reinforcement learning-based trajectory tracking optimal control for underactuated unmanned surface vehicles under asymmetric input saturation}, volume = {162}, year = {2025}, }
• (2025). Feature-enhanced edge-INverse attention network for skin lesion segmentation. EAAI, 162, 112306. (www)

  @article{ , abstract = {Skin cancer, particularly melanoma, is one of the most aggressive and deadly forms of cancer with its incidence rising globally. Early detection is crucial for improving survival rates, but the traditional dermatoscopy method is a highly time-consuming and subjective process. To resolve this issue, we propose a novel Feature-Enhanced Edge-INverse attention network (FEEINnet) model that helps to segment the skin lesion region more accurately. FEEINnet consists of three sub-networks: Feature Enhanced Mechanism (FEM) learns and extracts the fine-grained enhanced features from informative channels, the Edge Attention Mechanism (EAM) helps to precisely identify the edges of the lesion region and the INverse Attention Mechanism (INAM) generates inverse attention maps which emphasize the less confident or ambiguous regions thereby increasing the segmentation accuracy iteratively. These three sub-networks collectively help to improve feature extraction, enhance boundary detection, and refine segmentation maps, even in challenging scenarios with varying lesion sizes, shapes and pigmentation. FEEINnet consistently outperforms existing models, achieving a F1-score of 95.55%, 95.53%, and 94.52%; Intersection over Union (IoU) of 92.76%, 92.43%, and 91.34%; and Structural Similarity Index Measure (SSIM) of 94.63%, 93.51%, and 91.85% on the Human Against Machine 10000 (HAM10000), Pedro Hispano Hospital ( P H 2 ), and International Skin Imaging Collaboration 2018 (ISIC2018) datasets, respectively. The obtained results demonstrate that the proposed model has a greater ability to segment complex skin lesions more accurately.}, archive = {J_EAAI}, author = {Shivamm Warambhey and Aravindkumar Sekar}, doi = {10.1016/j.engappai.2025.112306}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112306}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Feature-enhanced edge-INverse attention network for skin lesion segmentation}, volume = {162}, year = {2025}, }
• (2025). A severity indicator for unsupervised fault severity shift detection of heating, ventilation, and air conditioning sub-system faults using a novel adaptive feature weighing method. EAAI, 162, 112305. (www)

  @article{ , abstract = {Fault-free functioning of Heating, Ventilation, and Air Conditioning (HVAC) systems is essential for reducing energy waste in modern-day buildings. Hence, data-driven approaches for HVAC fault detection have gained popularity. Faults become more severe with time. Fault detection reveals the presence of an anomaly, but it does not convey how critical the fault severity is. Fault severity indication provides this essential context, enabling urgent resource allocation to more severe faults, adding practical significance. However, faults being rare, obtaining substantial data at different severity levels to train supervised Machine Learning models is a realistic challenge. Therefore, we propose a method for estimating fault severity in an unsupervised setting. We define a robust Severity Indicator (SI) that reflects the shift in the severity levels of a fault. First, we define a healthy domain boundary for fault-free data using One-Class Support Vector Machines. SI scores are then computed using a novel adaptive feature weighing algorithm that assigns weights to individual features, adaptively, for every fault. We focus on detecting the shift in severity, rather than quantifying it. The study of the robustness of SI for different faults in HVAC subsystems, chillers, and air handling units (AHUs) yields consistently promising results. Our comparative analysis shows that our method outperforms the unweighted approach and existing state-of-the-art techniques for fault severity estimation. Notably, our method excels in detecting low-severity faults, addressing a common limitation in current methods.}, archive = {J_EAAI}, author = {Ramnath V. Prabhu Bam and Rajesh S. Prabhu Gaonkar and Clint Pazhayidam George}, doi = {10.1016/j.engappai.2025.112305}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112305}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A severity indicator for unsupervised fault severity shift detection of heating, ventilation, and air conditioning sub-system faults using a novel adaptive feature weighing method}, volume = {162}, year = {2025}, }
• (2025). An efficient anchor-free model for ore particle size detection. EAAI, 162, 112304. (www)

  @article{ , abstract = {Accurate detection of ore size is crucial in mineral processing, directly impacting equipment efficiency and product quality. However, traditional anchor-based models often struggle with the irregular shapes and varying scales of ore particles, resulting in limited performance. To overcome these challenges, an anchor-free detection framework was proposed. It incorporates a cross-stage partial bottleneck and a spatial pyramid pooling cross-stage partial connections (SPPCSCP-DualConv), both enhanced with dual convolution, to improve feature extraction and multi-scale fusion. In the backbone, the dual convolution module combines group convolution with heterogeneous convolution to improve feature diversity. The SPPCSCP-DualConv module further enhances feature representation in complex backgrounds. Additionally, a simplified path aggregation network (simPANet) feature fusion module is employed in the neck to refine the integration of multi-scale features. The proposed model was trained using a combination of binary cross-entropy, complete intersection over union (IoU), and distribution focal loss to optimize detection accuracy. The proposed model achieved a mean average precision of 86.80 % at an IoU threshold of .5 and 78.50 % across IoU thresholds from .5 to .95, surpassing existing methods while maintaining a lightweight architecture with only 10.10 million parameters and 89.45 giga floating point operations per second. Ablation studies confirmed the effectiveness of the simPANet and SPPCSPC-DualConv modules in enhancing feature representation. Generalization tests across mining sites with similar distributions demonstrated strong performance, although limitations remain for exceptionally large ore blocks due to dataset bias. The proposed model significantly improved the accuracy and efficiency of ore particle size detection, providing reliable real-time insights to improve grinding control and mineral processing operations.}, archive = {J_EAAI}, author = {Kanghui Zhang and Qingkai Wang and Guobin Zou and Jiawei Yang and Tao Song and Yang Liu and Daoxi Liu}, doi = {10.1016/j.engappai.2025.112304}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112304}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An efficient anchor-free model for ore particle size detection}, volume = {162}, year = {2025}, }
• (2025). Propeller fault-detection method for electric-propulsion aircraft using motor signals and generative model-based semi-supervised learning. EAAI, 162, 112301. (www)

  @article{ , abstract = {Failures in propulsion components, such as propellers, can critically affect flight safety; thus, early failure detection, preferably before flight, is essential. Traditional fault-diagnosis methods typically rely on additional sensors or operational data, which may not be available or practical in all situations. This study addresses these challenges by introducing motor-electric-signal-based fault diagnosis that is independent of airframe configuration and can detect faults, even when the aircraft is not in operation. However, difficulties arise owing to poor class variance in motor-electric-signal data and the challenge of obtaining fault data. To overcome these issues, a semi-supervised learning model based on a modified variational autoencoder-generative adversarial network (VAE-GAN) is proposed, which predicts faults using only normal motor-electric-signal data. Additionally, a new preprocessing method and patch-based ensemble inference technique are introduced to improve the poor class-variance characteristics of the data, thereby enhancing the prediction performance. This work demonstrates that propeller faults can be successfully diagnosed using motor-electric signals without the need for additional sensors or fault-data acquisition.}, archive = {J_EAAI}, author = {Sanga Lee and Dohyeong Kim and Minkyun Noh and Shinkyu Jeong and Jikang Kong and Youngjun Yoo}, doi = {10.1016/j.engappai.2025.112301}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112301}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Propeller fault-detection method for electric-propulsion aircraft using motor signals and generative model-based semi-supervised learning}, volume = {162}, year = {2025}, }
• (2025). A lightweight wood defect segmentation network via multi-dimension boundary perception and guidance. EAAI, 162, 112300. (www)

  @article{ , abstract = {Wood surface defect segmentation is extremely critical for defect refinement and quality control of wooden products. However, it is a challenging task to develop an efficient method with current algorithms due to the complicated characteristics of wood defects with obscure boundary, intraclass difference and interclass similarity. To address these issues, a lightweight network via multi-dimension boundary perception and guidance is proposed for precise segmentation of wood defects. At first, based on the Segformer, a boundary prediction branch is added to enrich detailed boundary information in the encoder, and supervised by the Gaussian signal and cosine similarity, to balance the effect of the boundary gradient information. Then, a double-flow enhancing module is designed to integrate the adjacent level features, by embedding two enhancing paths, to adaptively generate discriminative information of the defects. Finally, a binary segmentation head following the predicted map is introduced to strengthen the penalty for the false prediction results of the boundary. Experimental results demonstrate the proposed method outperforms the state-of-the-arts on our wood surface defect dataset, as well as on three public datasets.}, archive = {J_EAAI}, author = {Yuhang Zhu and Ye Lin and Zhezhuang Xu and Dan Chen and Kunxin Zheng and Yazhou Yuan}, doi = {10.1016/j.engappai.2025.112300}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112300}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight wood defect segmentation network via multi-dimension boundary perception and guidance}, volume = {162}, year = {2025}, }
• (2025). Sparse view tomographic reconstruction of elongated objects using learned primal-dual networks. EAAI, 162, 112295. (www)

  @article{ , abstract = {In the wood industry, logs are commonly quality screened by discrete X-ray scans on a moving conveyor belt from a few source positions. Typically, the measurements are obtained in a single two-dimensional (2D) plane (a “slice”) by a sequential scanning geometry. The data from each slice alone does not carry sufficient information for a three-dimensional tomographic reconstruction in which biological features of interest in the log are well preserved. In the present work, we propose a learned iterative reconstruction method based on the Learned Primal-Dual neural network, suited for sequential scanning geometries. Our method accumulates information between neighbouring slices, instead of only accounting for single slices during reconstruction. Evaluations were performed by training U-Nets on segmentation of knots (branches), which are crucial features in wood processing. Our quantitative and qualitative evaluations show that with as few as five source positions our method yields reconstructions of logs that are sufficiently accurate to identify biological features like knots (branches), heartwood and sapwood.}, archive = {J_EAAI}, author = {Buda Bajić and Johannes A.J. Huber and Benedikt Neyses and Linus Olofsson and Ozan Öktem}, doi = {10.1016/j.engappai.2025.112295}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112295}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Sparse view tomographic reconstruction of elongated objects using learned primal-dual networks}, volume = {162}, year = {2025}, }
• (2025). A deep Q-network-driven multi-objective evolutionary algorithm for distributed heterogeneous hybrid flow shop scheduling with worker fatigue. EAAI, 162, 112292. (www)

  @article{ , abstract = {The hybrid flow shop scheduling problem (HFSP) is a prominent challenge in advanced manufacturing systems. Existing research often overlooks the impact of workers in production shops or treats worker fatigue as a static parameter, failing to capture its nonlinear accumulation and recovery effects on processing efficiency. However, with the advent of Industry 5.0, there has been a growing emphasis on the critical role of human factors in production scheduling. As a result, designing an effective algorithm for HFSP that considers human factors has become a prominent research focus. In this paper, an extended distributed heterogeneous hybrid flow shop scheduling problem with the dynamic effects of worker fatigue (DHHFSP-WF) is investigated. To address this problem, a Deep Q-Network-based multi-objective optimization algorithm (DQNMOEA) is designed to minimize makespan, total energy consumption (TEC), and total worker idle time (WIT). In DQNMOEA, a four-dimensional vector encoding scheme considering worker allocation represents solution, and a reconstruction strategy ensures initial population quality and diversity. Moreover, an improved order crossover, two-point crossover, and a segment-based recombination mutation method are proposed to enhance the global search performance of the algorithm. Then, a problem-specific local search strategy is designed for each layer of the vector, allowing the Deep Q-Network (DQN)-based adaptive decision-making mechanism to perform local perturbations on the current non-dominated solutions in the most suitable dimensions. Finally, seven algorithms are adopted to make a comparison on 36 sets of instances, the experimental results indicate that DQNMOEA exhibits competitive performance in solving DHHFSP-WF.}, archive = {J_EAAI}, author = {Jianlin Zhang and Longbin Ma and Wu Zhao and Jie Cao and Zuohan Chen and Tianpeng Xu}, doi = {10.1016/j.engappai.2025.112292}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112292}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A deep Q-network-driven multi-objective evolutionary algorithm for distributed heterogeneous hybrid flow shop scheduling with worker fatigue}, volume = {162}, year = {2025}, }
• (2025). An angle-enhanced deep learning framework for thermal defect diagnosis in overhead transmission line composite insulators. EAAI, 162, 112285. (www)

  @article{ , abstract = {Due to long-term outdoor exposure, composite insulators are susceptible to degradation and abnormal temperature rise, making Unmanned Aerial Vehicle (UAV)-based infrared inspections essential for effective monitoring. However, traditional manual interpretation of these images is inefficient and subjective. To improve detection automation and accuracy, we propose an intelligent detection method for composite insulators in infrared images based on an improved You Only Look Once version 11 (YOLOv11) model. The proposed approach introduces Oriented Bounding Boxes (OBBs) for annotation and designs an Angle-Enhanced Probabilistic Intersection over Union (AE-ProbIoU) loss function to enhance the model's ability to detect rotated objects. Experimental results demonstrate that the proposed Angle-Enhanced You Only Look Once (AE-YOLO) model achieves a mAP50:95 of 94.0 % and an angle prediction accuracy of 94.3 %. In addition, a temperature extraction module based on the OBBs is developed to accurately derive the temperature profile of the insulator core rod. This method significantly enhances the intelligence level of infrared image analysis for composite insulators and provides technical support for condition assessment and fault prediction in power transmission lines.}, archive = {J_EAAI}, author = {Xinzhe Yu and Zhenan Zhou and Yu Deng and Kun Zhang and Chen Gu and Zheyuan Liu and Songsong Zhou}, doi = {10.1016/j.engappai.2025.112285}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112285}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An angle-enhanced deep learning framework for thermal defect diagnosis in overhead transmission line composite insulators}, volume = {162}, year = {2025}, }
• (2025). Bayesian bidirectional long short-term memory-based kinematics-dynamics fusion for fault-tolerant vehicle state estimation under yaw rate sensor failures. EAAI, 162, 112283. (www)

  @article{ , abstract = {Accurate estimation of vehicle states is a fundamental component of vehicle stability control systems. To address the issue of inaccurate estimation of vehicle state parameters resulting from yaw rate sensor failures, this study proposes a three-mode collaborative fault-tolerant state estimation method based on Bayesian Bidirectional Long Short-Term Memory (BiLSTM) kinematics-dynamics fusion. First, the kinematics-based method is established using the kinematics model. Second, the dynamics-based method is designed by integrating the Unscented Kalman Filter (UKF) with the dynamics model. Subsequently, a BiLSTM network fusion model based on Bayesian optimization is presented. The model utilizes estimates from kinematic and kinetic methods as a priori inputs and combines the bidirectional information capturing capability of BiLSTM with hyperparameter tuning from Bayesian optimization. The results indicate that when the yaw rate sensor fails, the proposed method achieves an average Root Mean Square Error (RMSE) of 0.0276 km per hour (km/h) for longitudinal speed, 0.0008 radian (rad) for side slip angle, and 0.0072 radian per second (rad/s) for yaw rate across all scenarios. This performance demonstrates a superiority over various maneuvers. This paper combines kinematics, dynamics, and deep learning to provide a reliable solution for fault-tolerant estimation of vehicle states.}, archive = {J_EAAI}, author = {Min Gao and Jiaqi Li and Wei Wang and Renguang Wang and Jin Luo and Jing Li}, doi = {10.1016/j.engappai.2025.112283}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112283}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Bayesian bidirectional long short-term memory-based kinematics-dynamics fusion for fault-tolerant vehicle state estimation under yaw rate sensor failures}, volume = {162}, year = {2025}, }
• (2025). Hierarchical binary diffing with inductive code representation learning with graph sampling-and-aggregate. EAAI, 162, 112279. (www)

  @article{ , abstract = {Binary program diffing, or simply binary diffing, is a type of program analysis technique that quantifies the similarity between two binary programs to derive their differences. In particular, binary diffing is an essential technique for uncovering vulnerabilities and potential attack vectors in industrial control systems, where patch deployment is complicated by closed and restricted environments. Studies on binary diffing can be broadly categorized into dynamic analysis-based, static analysis-based, and neural network-based approaches. Each category of existing studies has its shortcomings, including limited coverage, low accuracy, and issues with on-demand learning. In this paper, we propose the binary diffing with sampling-and-aggregate, a hierarchical binary diffing model that generates inductive code representations based on graph sampling-and-aggregate. Our model sequentially produces instruction-level embedding, block-level embedding, and function-level embedding from the inter-procedural control flow graph of a given program, and then performs hierarchical code diffing based on these embeddings. We formally define the detailed models and present the algorithm of hierarchical binary diffing. Additionally, we conduct a thorough analysis of this algorithm, deriving several advantages. We implemented a prototype and evaluated it on a large-scale dataset in a cross-version, cross-optimization, and obfuscation settings. Our prototype showed F1-scores up to 0.96 and 0.968 in cross-version setting for function and basic block diffing, respectively. Also, our method demonstrated its robustness over several binary obfuscations. In conclusion, our proposal, which generates basic block- and function-level embedding by considering the control flow, has solid advantages on binary diffing and shows the robustness on the binary tampering.}, archive = {J_EAAI}, author = {Seungho Jeon and Kijong Koo and Daesung Moon and Jung Taek Seo}, doi = {10.1016/j.engappai.2025.112279}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112279}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hierarchical binary diffing with inductive code representation learning with graph sampling-and-aggregate}, volume = {162}, year = {2025}, }
• (2025). Nonlinear excavation load prediction of hydraulic excavator based on gated recurrent unit neural network. EAAI, 162, 112278. (www)

  @article{ , abstract = {Excavator arms are integral to the mining and construction industries, where real-time excavation load prediction is a critical element for the advancement of automated excavation technology. This study presents a novel Physics-guided Neural Network (PGNN) designed to predict the excavating force of hydraulic cylinders used in earthwork excavation. The PGNN model synergizes the physical load model of excavators with a Gated Recurrent Unit (GRU) neural network and is optimized using the Hyperband algorithm to attain both high-speed and precise forecasting. Through comparative experiments, the study validates the PGNN model's ability to achieve optimal response speed and precision in predicting excavation loads. Additionally, the predictive performance of the PGNN model is assessed via a Hardware-in-the-loop (HIL) test, conducted within the context of an actual excavation experiment. This research introduces a promising approach that seamlessly integrates physics-based modeling with machine learning techniques, facilitating real-time load forecasting for excavators. The findings pave the way for more efficient and precise excavation processes, with implications for the broader fields of mining and construction automation.}, archive = {J_EAAI}, author = {Jinshi Chen and Yue Yu and Dongyang Huo and Han Zhang and Jingyan Wang}, doi = {10.1016/j.engappai.2025.112278}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112278}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Nonlinear excavation load prediction of hydraulic excavator based on gated recurrent unit neural network}, volume = {162}, year = {2025}, }
• (2025). Plug-and-play fine grained neural cognitive diagnosis framework. EAAI, 162, 112276. (www)

  @article{ , abstract = {Cognitive diagnosis (CD) is a core task in intelligent education, which accurately assesses students’ mastery of specific knowledge concepts (KCs) by analyzing their answer records. However, existing methods mainly rely on explicit interaction data and use diagnostic models for automatic knowledge proficiency inference. These methods lack systematic optimization for fine-grained knowledge level representation, making it difficult to fully reflect students’ true learning status. To address this, this paper introduces a Plug-and-Play F ine Grained N eural C ognitive D iagnosis Framework (FNCD) with Knowledge-Level Constraint Awareness . The framework combines a knowledge proficiency evaluation module with students’ answer records and a Q-matrix to statically assess knowledge mastery. It uses a student similarity construction method based on random grouping to reveal latent learning pattern associations. Additionally, it employs a multi-scale relational learning strategy and a Top-k attention-enhanced graph network mechanism to dynamically adjust the student similarity relationship network, accurately modeling the complex learning relationships between students. Ultimately, a joint training mechanism is used to optimize the outputs of each module, significantly improving the rationality, interpretability, and accuracy of CD. The experimental results demonstrate that FNCD, as an artificial intelligence-driven plug-and-play module, can be effectively integrated into existing CD models to enhance the modeling of fine-grained knowledge mastery and improve diagnostic accuracy, showcasing the application potential of artificial intelligence in personalized education.}, archive = {J_EAAI}, author = {Xinjie Sun and Qi Liu and Kai Zhang and Weiyin Gong and Shuanghong Shen and Fei Wang and Yan Zhuang and Meikai Bao and Shijin Wang and Yuling Ma and Enhong Chen}, doi = {10.1016/j.engappai.2025.112276}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112276}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Plug-and-play fine grained neural cognitive diagnosis framework}, volume = {162}, year = {2025}, }
• (2025). Open-vocabulary object detection via neighboring region attention alignment. EAAI, 162, 112270. (www)

  @article{ , abstract = {The nature of diversity in real-world environments necessitates neural network models to expand from closed category settings to accommodate novel emerging categories. In this paper, we study open-vocabulary object detection (OVD), which facilitates the detection of novel object classes under the supervision of only base annotations and open-vocabulary knowledge. However, we find that the inadequacy of distilled information from the detector head during the alignment process inevitably constrains the performance of recent distillation-based OVD strategies. To this end, we propose Neighboring Region Attention Alignment (NRAA), which performs alignment within the attention mechanism to boost the open-vocabulary inference. Specifically, for a given proposal region, we randomly explore the neighboring boxes to capture the surrounding contextual vocabulary knowledge. Then, a set of regional token features, encompassing both the proposal and neighboring regions, utilize our proposed Neighboring Region Attention (NRA) to extract interaction information. Finally, this information is seamlessly provided to the distillation procedure to assist the alignment between the detector and the pre-trained vision-language models (VLMs). Extensive experiments validate that our proposed model exhibits superior performance on open-vocabulary benchmarks.}, archive = {J_EAAI}, author = {Sunyuan Qiang and Xianfei Li and Yanyan Liang and Wenlong Liao and Tao He and Pai Peng}, doi = {10.1016/j.engappai.2025.112270}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112270}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Open-vocabulary object detection via neighboring region attention alignment}, volume = {162}, year = {2025}, }
• (2025). A time and frequency convolutional autoencoder for anomaly detection in industrial robots based on inertial measurement unit error calibration. EAAI, 162, 112269. (www)

  @article{ , abstract = {In the realm of industrial robots, ensuring operational reliability and Long-Term Autonomy hinges on the accurate detection of anomalies. However, this sample difference due to noise, joint random errors and sensor errors increases the challenge of robot anomaly detection. To address this problem, an unsupervised deep learning method based on inertial measurement unit (IMU) error calibration is proposed. Firstly, the attitude signals acquired by the IMU from the end of the robot were calibrated using Kalman filtering. The three dimensional (3D) free acceleration was corrected based on the calibrated attitude signal and the calibrated 3D free acceleration signal was used as a signal sample. Secondly, a time and frequency convolutional autoencoder model (TFCAE) is proposed. And the distribution of the different component signals is fitted by stacking multiple encoder modules and 3D-TFCAE is used for 3D free acceleration signal reconstruction model. Then, the error sphere radius is calculated based on the reconstruction error of the 3D free acceleration signal. And the error sphere radius is used as the anomaly detection threshold to realize the robust detection of different types of anomalies. The model was evaluated on a constructed anomaly dataset. This study contributes an innovative 3D-TFCAE architecture, integrating Kalman filtering with time-frequency feature fusion, markedly enhancing anomaly detection in complex signal environments. Experimental findings reveal that 3D-TFCAE significantly outperforms 18 baseline models, improving detection accuracy by about 20 %–40 %, offering an effective solution for high-precision anomaly detection in industrial robots. The code for this project is available at https://github.com/LJlong977/3DTFCAE .}, archive = {J_EAAI}, author = {Jianlong Li and Xiaoqin Liu and Xing Wu and Dongxiao Wang and Kai Xu and Yashan Li}, doi = {10.1016/j.engappai.2025.112269}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112269}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A time and frequency convolutional autoencoder for anomaly detection in industrial robots based on inertial measurement unit error calibration}, volume = {162}, year = {2025}, }
• (2025). A lightweight citrus detection and counting method based on deep learning model. EAAI, 162, 112268. (www)

  @article{ , abstract = {Picking robots have become an important development direction of smart agriculture, and the accurate detection, counting and lightweight deployment of fruits are the technical basis for realizing robot picking. However, due to complex weather conditions and the possible mutual occlusion between branches and citrus, it is challenging to accurately detect and count citrus in orchards. This study proposes a lightweight small target detection model for detecting and counting citrus, and deploys it on the citrus detection platform. The model first introduces FasterNet Block into the cross-stage partial feature fusion module of the backbone network to reduce the number of parameters and calculations while improving the detection accuracy of the network. Secondly, a multi-scale attention mechanism is added to the backbone network to enhance the feature extraction ability of the network. Finally, a bounding box loss function based on a dynamic non-monotonic focusing mechanism is used to increase the model convergence speed and further improve the model accuracy. Experimental results show that the model has an accuracy of 92.7%, an average precision of 91.7%, and a model size of only 5.37 megabytes. The lightweight model is applied to the citrus detection platform. Based on this application, a citrus counting method is proposed, which obtains a mean absolute error (MAE) of 0.92, a root mean square error (RMSE) of 1.28, a determination coefficient ( R 2 ) of 0.98, and a frame rate of 80.6 per second, which meets the requirements of real-time citrus detection and counting. This provides technical support for the subsequent deployment and counting research of picking robots.}, archive = {J_EAAI}, author = {Jiqing Chen and Mingchang Zhang and Bin Lu and Quan Chen and Zhiwu Jiang and Peilin Li and Jingyao Gai}, doi = {10.1016/j.engappai.2025.112268}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112268}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight citrus detection and counting method based on deep learning model}, volume = {162}, year = {2025}, }
• (2025). Medical image fusion for high-level analysis: A mutual enhancement framework for unaligned photoacoustic tomography and magnetic resonance imaging. EAAI, 162, 112259. (www)

  @article{ , abstract = {Photoacoustic tomography (PAT) offers optical contrast, whereas magnetic resonance imaging (MRI) excels in imaging soft tissue and organ anatomy. The fusion of PAT with MRI holds promising application prospects due to their complementary advantages. Existing image fusion has made considerable progress in pre-registered images, yet spatial deformations are difficult to avoid in medical imaging scenarios. More importantly, current algorithms focus on visual quality and statistical metrics, thus overlooking the requirements of high-level tasks. To address these challenges, we propose an unsupervised fusion model, termed PAMRFuse+, which integrates image generation and registration. Specifically, a cross-modal style transfer network is introduced to simplify cross-modal registration to single-modal registration. Subsequently, a multi-level registration network is employed to predict displacement vector fields. Furthermore, a dual-branch feature decomposition fusion network is proposed to address the challenges of cross-modal feature modeling and decomposition by integrating modality-specific and modality-shared features. PAMRFuse+ achieves satisfactory results in registering and fusing unaligned PAT–MRI datasets. Moreover, for the first time, we evaluate the performance of medical image fusion with multi-organ instance segmentation. Extensive experimental demonstrations reveal the advantages of PAMRFuse+ in improving the performance of medical image analysis tasks.}, archive = {J_EAAI}, author = {Yutian Zhong and Jinchuan He and Zhichao Liang and Shuangyang Zhang and Qianjin Feng and Lijun Lu and Li Qi}, doi = {10.1016/j.engappai.2025.112259}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112259}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Medical image fusion for high-level analysis: A mutual enhancement framework for unaligned photoacoustic tomography and magnetic resonance imaging}, volume = {162}, year = {2025}, }
• (2025). Max–Min pooling and squeeze excitation lightweight bidirectional mamba for image classification. EAAI, 162, 112246. (www)

  @article{ , abstract = {To address challenges stemming from the quadratic computational complexity, slow execution speed, and high memory consumption of Transformer models, we propose a novel image classification method: Max–Min Pooling and Squeeze Excitation Lightweight Bidirectional Mamba(MMPSELMamba). The core artificial intelligence innovations include: (1) We design a max–min pooling mechanism that synergistically preserves high-activation foreground features through max-pooling and low-intensity contextual details via min-pooling. This approach addresses the information loss problem in conventional single-mode pooling methods; (2) Inspired by squeeze-enhanced Axial Transformer(SeaFormer), we design an axial squeeze excitation module that compresses redundant features along both vertical and horizontal dimensions while enhancing discriminative feature refinement; (3) Building upon recent advances in sequence modeling, we replace self-attention with a bidirectional Mamba architecture based on state space models (SSM), achieving linear complexity in the model of long-range dependencies; (4) Our novel multi-scale integration unit combines upsampling, concatenation, and downsampling operations to optimize feature fusion while minimizing computational overhead. For engineering applications, MMPSELMamba is specifically designed for resource-constrained environments such as edge devices and mobile vision systems. By integrating depthwise separable convolutions and lightweight SSM operations, it achieves a 36% parameter reduction and a 16% lower computational load compared to SeaFormer, while maintaining competitive accuracy. Experiments on public datasets validate its deployment potential in real-world scenarios like autonomous drones and embedded surveillance. Confusion matrices and heatmap visualizations further confirm the superior performance of the MMPSELMamba method. Our code can be gained from: https://github.com/yalemitter/MMPSELMamba-main .}, archive = {J_EAAI}, author = {Senlin Chi and Zhiwen Wang and Lianyuan Jang and Mengsi Gong}, doi = {10.1016/j.engappai.2025.112246}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112246}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Max–Min pooling and squeeze excitation lightweight bidirectional mamba for image classification}, volume = {162}, year = {2025}, }
• (2025). Stacked machine learning-based probabilistic seismic demand and risk assessment of railway embankments with variable slopes. EAAI, 162, 112245. (www)

  @article{ , abstract = {Accurate seismic risk assessment of railway embankments is critical for risk mitigation, seismic design, and emergency planning. However, conventional methods often suffer from computational inefficiency and complexity. This study proposes a novel machine learning (ML) framework to rapidly and accurately evaluate probabilistic seismic demand and risk for railway embankments. Latin hypercube sampling is utilised to generate representative soil parameter samples to construct numerical models for simulating dynamic responses under near-fault pulse-like ground motions. The peak permanent settlement (PPS) of the embankment surface is used as the key performance metric. Multiple ML models, including decision trees, random forests (RFs), extreme gradient boosting (XGBoost), artificial neural networks (ANNs), and a stacked ML model that integrates RFs, XGBoost, and ANNs, are trained and compared. The stacked ML model outperforms the other models and achieves the highest predictive accuracy for the PPS. SHapley Additive exPlanations are used to identify the velocity spectrum intensity (VSI) and the internal friction angle of the embankment as the most influential factors. Seismic fragility and risk curves are subsequently developed. The VSI and a power-law seismic hazard function are combined to estimate the annual exceedance probabilities for three seismic design criteria levels. The proposed ML framework significantly enhances the efficiency of seismic risk analysis while maintaining high precision, thereby providing a transformative approach for the seismic assessment of railway embankments.}, archive = {J_EAAI}, author = {Pan Si and Liang Tang and Shuang Tian and Xianzhang Ling and Yanfang Liu}, doi = {10.1016/j.engappai.2025.112245}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112245}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Stacked machine learning-based probabilistic seismic demand and risk assessment of railway embankments with variable slopes}, volume = {162}, year = {2025}, }
• (2025). Quantization-based deep diversified ensemble for medical image segmentation. EAAI, 162, 112242. (www)

  @article{ , abstract = {Recent advancements in fully convolutional networks (FCNs) have significantly improved medical image segmentation. Ensemble methods are often used to further enhance performance, with diversity among learners being a critical factor. However, many current approaches focus on diversifying training samples or predictions while overlooking the diversity of internal multi-scale features. This oversight can lead to high correlations among features across different learners, limiting overall effectiveness. Additionally, traditional quantization methods aim to minimize accuracy loss by maintaining a rigid quantization process. This rigidity can eliminate the randomness introduced by quantization, further reducing ensemble diversity and effectiveness. In this paper, we propose a novel approach called Quantization-based Deep Diversified Ensemble (QDD-Ens) for medical image segmentation. Our method enhances the diversity of internal features among ensemble learners through two mechanisms: deep diversified loss, which focuses on feature diversity rather than segmentation accuracy, and deep diversified quantization, which preserves beneficial randomness in quantization process. Furthermore, QDD-Ens facilitates a deeper form of ensemble learning by employing a meta-learner to integrate diversified features at multiple resolution levels from various base learners, which are diversified by two above diversify enhancement mechanisms. Extensive experiments on five public medical image segmentation datasets show that our method significantly improves segmentation accuracy and outperforms existing ensemble techniques. The source code is publicly available to support future research. ( https://github.com/JerRuy/QDD-Ens )}, archive = {J_EAAI}, author = {Jiawei Zhang and Jialin Wang and Qi Wang and Yanchun Zhang and Weihong Han and Yangyang Mei and Yiyu Shi and Jian Zhuang and Meiping Huang and Xiaowei Xu}, doi = {10.1016/j.engappai.2025.112242}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112242}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Quantization-based deep diversified ensemble for medical image segmentation}, volume = {162}, year = {2025}, }
• (2025). Machine learning methods comparison by using statistical tests in solar energy forecasting based on weather features. EAAI, 162, 112239. (www)

  @article{ , abstract = {Climate change necessitates precise solar forecasting due to its weather-dependent intermittency. Key parameters - temperature, visibility, altitude, pressure, and wind speed - were analyzed using non-parametric tests. We prioritized short-term weather patterns over random data splitting for enhanced accuracy.Non-parametric tests, such as the Kolmogorov-Smirnov test, were used to assess data normality and select highly correlated features. Principal Component Analysis (PCA) reduces dataset dimensionality while preserving critical trends. Various machine learning approaches were evaluated, including: weighted linear regression (both with and without dimensionality reduction), boosted regression trees, and deep learning architectures-comprising both fundamental models (Convolutional Neural Networks [CNNs] and Recurrent Neural Networks [RNNs]) and advanced hybrid architectures (Temporal Convolutional Networks (TCN) Convolutional Neural Network-Long Short-Term Memory network (CNN-LSTM). All models were optimized through systematic hyperparameter tuning to enhance predictive performance, reduce computational complexity, and improve learning convergence rates. Special attention was given to addressing vanishing gradient problems in deep neural network implementations. Results show TCN outperform other deep learning models, achieving lower training and testing errors with fewer parameters and reduced time complexity. CNN-LSTM models, designed for spatial-sequence prediction, perform well but require more parameters and computational time. The lowest test and training errors belong to CNN-LSTM and TCN, with approximately 9 % and 2 % lower than the maximum amount, respectively. A trade-off between model complexity, error rates, and computational efficiency must be considered when selecting the optimal approach. Since relevant weather features vary by location, the proposed methodology serves as an adaptable algorithm for solar energy prediction in diverse geographical regions.}, archive = {J_EAAI}, author = {Mohammadreza pourmir and Seyedeh Mohadeseh Miri}, doi = {10.1016/j.engappai.2025.112239}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112239}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Machine learning methods comparison by using statistical tests in solar energy forecasting based on weather features}, volume = {162}, year = {2025}, }
• (2025). Adaptive chemical industrial processes fault detection model based on sparse filtering-based improved mixed-gaussian probabilistic principal component analysis considering low-probability events. EAAI, 162, 112236. (www)

  @article{ , abstract = {Probabilistic Principal Component Analysis (PPCA) is widely used in process monitoring. However, its underlying assumption that data follows a Gaussian distribution limits its effectiveness in handling Low Probability Events (LPEs), which often deviate from this assumption. To address this challenge, we propose a novel method called Sparse Filtering-based Improved Mixed-Gaussian Probabilistic Principal Component Analysis (SFIMPPCA) for enhanced LPEs detection. First, a Sparse Filtering (SF) preprocessing technique with an incremental structure is employed to extract the most discriminative features. Second, to address the distortion caused by LPEs, a dynamic ratio correction mechanism based on statistical variability is introduced, followed by a newly designed Mixed-Gaussian Probabilistic Principal Component Analysis (MPPCA). Third, a Bayesian Optimization Algorithm (BOA) is applied to automatically adjust control limits, enhancing the accuracy and reliability of fault detection. The effectiveness of the proposed method is validated using the Tennessee Eastman (TE) process and the Tin Chemical Process (TCP). Experimental results demonstrate that the proposed method significantly improves performance under LPEs conditions, achieving a 10%–12% improvement in most cases.}, archive = {J_EAAI}, author = {Chuangyan Yang and Jiande Wu and Peng Li and Xun Lang and Mingxi Ai and Hancheng Wang}, doi = {10.1016/j.engappai.2025.112236}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112236}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive chemical industrial processes fault detection model based on sparse filtering-based improved mixed-gaussian probabilistic principal component analysis considering low-probability events}, volume = {162}, year = {2025}, }
• (2025). Dynamic graph neural network with historic-sequential information representation. EAAI, 162, 112234. (www)

  @article{ , abstract = {In today’s interconnected world, networks, whether social, technological, or biological, are constantly evolving, with relationships forming, shifting, and dissolving over time. Traditional models struggle to capture this fluidity, treating networks as static snapshots rather than living systems. While existing research has made strides in analyzing either spatial–temporal patterns or temporal sequences separately, these approaches often fall short in scalability and fail to unify discrete and continuous-time perspectives. To bridge this gap, we introduce DyGHS (Dynamic Graph Historical-Sequential), an innovative framework that harnesses dynamic graph neural networks to model how nodes, edges, and their historical interactions evolve together. By seamlessly integrating discrete-time and continuous-time graph representations, DyGHS not only captures the richness of real-world networks but also efficiently predicts future connections and node behaviors. Our experiments reveal that combining continuous-time Fourier transform (CTFT) with graph neural networks significantly boosts prediction accuracy, outperforming current methods in tasks like link prediction and node classification. This advancement opens new doors for understanding and anticipating the ever-changing tapestry of networked systems.}, archive = {J_EAAI}, author = {Adam Abakar Hamid and Anping Zhao and Alladoumbaye Ngueilbaye}, doi = {10.1016/j.engappai.2025.112234}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112234}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dynamic graph neural network with historic-sequential information representation}, volume = {162}, year = {2025}, }
• (2025). MAE-diff: Masked-AutoEncoder-guided diffusion framework for source-free domain adaptive medical image segmentation. EAAI, 162, 112219. (www)

  @article{ , abstract = {Domain gaps can often cause dramatic performance deterioration when applying medical image segmentation models trained on the source domain to the target domain. Although unsupervised domain adaptation methods can address the domain gap challenge to some extent, their reliance on accessing source images largely hampers their practical applicability, as source data are often inaccessible due to privacy concerns. Moreover, the low-quality characteristic of medical images can further degrade the domain adaptation performance of segmentation models. To address these issues, we propose the Masked-AutoEncoder-guided Diffusion (MAE-Diff) framework for source-free domain adaptive medical image segmentation. MAE-Diff mainly consists of a Masked AutoEncoder (MAE) Module for effective feature extraction and domain adaptation, and a Diffusion Module for effective segmentation of low-quality medical images. On source images, the MAE encoder is trained to extract image-specific features, and the Diffusion Module is trained to generate segmentation maps following a gradual denoising strategy, under the guidance of features extracted by the MAE encoder. Training on the target domain involves only fine-tuning MAE (trained on the source images) with target images, allowing MAE-Diff to adapt to the target domain distribution. Inference on target images can then be made by the source-based Diffusion Module, under the guidance of features extracted by the MAE encoder fine-tuned on the target images. Extensive experiments on three datasets demonstrate the effectiveness of the proposed framework for source-free domain-adaptive medical image segmentation. The code of MAE-Diff is available at https://github.com/xuss804/MAEDiff .}, archive = {J_EAAI}, author = {Shanshan Xu and Le Xu and Yeqing Yang and Lixia Tian}, doi = {10.1016/j.engappai.2025.112219}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112219}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {MAE-diff: Masked-AutoEncoder-guided diffusion framework for source-free domain adaptive medical image segmentation}, volume = {162}, year = {2025}, }
• (2025). A hierarchical reinforcement learning framework with imitation learning and bayesian Actor–Critic for distributed unmanned underwater vehicle encirclement in dynamic maritime environments. EAAI, 162, 112214. (www)

  @article{ , abstract = {The rapid development of Unmanned Underwater Vehicles (UUVs) has introduced potential threats to critical infrastructure, with existing methods facing challenges in multi-agent collaboration and dynamic environment adaptation. To address these challenges, we propose HRL-IL-BAC—a Hierarchical Reinforcement Learning framework equipped with an Imitation-Learning (IL) high-level policy and a Bayesian Actor–Critic (BAC) low-level controller. The framework decomposes the encirclement mission into four sequential phases: submergence, interception, encirclement, and standoff tracking. At the high level, multi-agent collaboration and long-horizon autonomous decision-making are orchestrated by the IL-guided policy, whereby strategy convergence is accelerated and phase selection is rendered more precise. At the low level, environmental and model uncertainties are explicitly modeled by the BAC controller, so that adaptability and control accuracy are markedly boosted under dynamic ocean conditions. Validation through simulations and physical experiments demonstrates that the HRL-IL-BAC framework achieves 23.6% faster encirclement efficiency and 52.0% enhanced system stability compared to conventional methods, while maintaining superior adaptability to dynamic maritime conditions. This work provides an innovative and effective solution for underwater defense and marine security, offering new insights into multi-agent systems for complex maritime operations.}, archive = {J_EAAI}, author = {Hanbin Zhang and Wenchuan Zang and Peng Yao and Jiangli Cao and Dalei Song}, doi = {10.1016/j.engappai.2025.112214}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112214}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A hierarchical reinforcement learning framework with imitation learning and bayesian Actor–Critic for distributed unmanned underwater vehicle encirclement in dynamic maritime environments}, volume = {162}, year = {2025}, }
• (2025). Dynamic-static siamese Takagi–Sugeno-kang fuzzy system with inductive-reflection deep fuzzy rule. EAAI, 162, 112199. (www)

  @article{ , abstract = {The first-order Takagi–Sugeno–Kang (TSK) fuzzy classifiers are famous for their high computational efficiency and strong interpretability, but they often struggle to learn from complex and large-scale datasets, and perform not very well compared to higher-order TSK fuzzy classifiers. To address this issue, in this paper we propose a novel Dynamic-Static Siamese TSK Fuzzy classifier with Inductive-Reflection Deep Fuzzy rule. It aims to enhance the model’s self-learning capabilities by utilizing Siamese network to integrate deep fuzzy knowledge and fine-grained knowledge without the need for a teacher model. The innovations of this study are as follows: (1) The deep fuzzy rules in the proposed classifier are enriched with an “Inductive-Reflection” process, which reduces constraints on traditional basic fuzzy rule and aligns rule acquisition more closely with general human thinking manners; (2) The proposed method includes a mechanism for self-learning and improvement from both deep and fine-grained fuzzy knowledge, eliminating the complexity of retraining a new teacher model; (3) An adaptive learning function is developed to effectively adjust the learning process, adapting to tasks with different complexities. Extensive experiments results on benchmark datasets, as well as two real-world datasets, demonstrate the effectiveness of the proposed classifier in terms of classification accuracy and weighted F1-score.}, archive = {J_EAAI}, author = {Xiongtao Zhang and Qihuan Shi and Yunliang Jiang and Qing Shen and Jungang Lou and Ruiqin Wang}, doi = {10.1016/j.engappai.2025.112199}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112199}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dynamic-static siamese Takagi–Sugeno-kang fuzzy system with inductive-reflection deep fuzzy rule}, volume = {162}, year = {2025}, }
• (2025). Enhancing anomaly detection with few-shot fine-tuned long text-to-image models. EAAI, 162, 112174. (www)

  @article{ , abstract = {Industrial anomaly detection plays a crucial role in the industrial manufacturing field. Currently, utilizing generated data to improve the performance of the anomaly detection model is an effective approach. However, most existing methods often rely on mask-guided synthesis, where the distribution of the generated defects is limited by masks that are typically random or learned by a model. In addition, the scarcity of real anomalous samples makes it difficult for generative models to capture genuine defect patterns and align with the real anomaly distribution. To tackle these issues, we propose DefectGen, the first long-text-guided few-shot text-to-image data generation pipeline for industrial anomaly detection. To improve distribution alignment under limited anomaly samples, DefectGen incorporates a Prompt Generation and Variation Module, which uses MLLMs (Multimodal Large Language Models) to expand few-shot image–text pairs into diverse and semantically rich prompts, and DoKr (Weight- D ecomposed L o w-Rank Adaptation with Kr onecker product), a lightweight fine-tuning strategy with structured low-rank adaptation. To ensure the quality of synthetic data, DefectGen further introduces the Real-Guided Clustering Filter, which selects high-quality generated samples by comparing their features with those of real anomalies. Experiments on the MVTec AD(MVTec AnomalyDetection) dataset show that DefectGen generates more diverse and realistic synthetic anomalies and achieves a 5.58% average improvement in anomaly classification accuracy compared to state-of-the-art methods. Code and data are available at: https://anonymous.4open.science/r/DefectGen-CD04/ .}, archive = {J_EAAI}, author = {Jiachen Liu and Jiajia An and Junbin Lu and Zhuoqin Yang and Jinbao Wang and Ping Lu and Yuying Wang and Linlin Shen}, doi = {10.1016/j.engappai.2025.112174}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112174}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing anomaly detection with few-shot fine-tuned long text-to-image models}, volume = {162}, year = {2025}, }
• (2025). A database of dentition images of indian breed cattle and estimation of cattle’s age using deep learning algorithms. EAAI, 162, 112172. (www)

  @article{ , abstract = {Accurate age estimation in livestock is crucial in agricultural management, influencing breeding, healthcare, and production planning. The primary method entails visually inspecting the teeth, a procedure susceptible to human fallibility. Recent deep learning (DL) algorithms are highly efficient in automating age prediction based on dentition images, offering faster and more accurate results. To our knowledge, no dataset exists for Indian livestock, hindering DL algorithm development for teeth segmentation and age estimation. Therefore, we created a database of dentition images for Indian livestock, covering an age range of 1 to 15 years, using 2,883 subjects. In this research, we employed a hybrid model based on You Only Look Once version 8 (YOLOv8) for instance segmentation of raw dentition images. Segmentation achieved 73.59% mean Average Precision (mAP) 50 and 56.44% mAP50–90, indicating strong performance in varying localization thresholds. These segmented images were then used in various DL algorithms, including convolutional neural network (CNN)-based transformers and vision transformers (ViTs), to develop DL models for age estimation. We conducted three age estimation experiments with increasing granularity: four age groups (4-year intervals), seven age groups (2-year intervals), and fifteen age groups (1-year intervals). InceptionV3 achieved the highest accuracy in the first two experiments with 73.72% and 62.41%, respectively. In the third experiment, MobileNet performed best with an accuracy of 43.87%. Our results demonstrate the effectiveness of these intricate DL models in enhancing livestock age estimation, making the process more streamlined and reliable. The study may assist farmers in making informed and efficient agricultural management decisions.}, archive = {J_EAAI}, author = {Chinmay Vijay Patil and Ankit Ashokrao Bhurane and Preeti Ghasad and Vipin Kamble and Manish Sharma and Anand Singh and Nareshkumar Nandeshwar and Ru-San Tan and Rajendra Acharya}, doi = {10.1016/j.engappai.2025.112172}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112172}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A database of dentition images of indian breed cattle and estimation of cattle’s age using deep learning algorithms}, volume = {162}, year = {2025}, }
• (2025). Physics-informed fine-tuning for physics discovery from random and sparse data. EAAI, 162, 112132. (www)

  @article{ , abstract = {Although significant advances have been made in both numerical methods and machine learning approaches for solving differential equations across various scientific sectors, these methods often rely on complete information about the differential equations, including precise parameter values, which are not easily obtainable in real-world scenarios. To address this challenge, data-driven methods for discovering differential equations have gained growing popularity in recent years. However, many existing approaches demand unrealistic prerequisites, such as extensive high-fidelity data or carefully designed low-fidelity data and functions. In this paper, we propose a novel method: P hysics- I nformed F ine- T uning ( PIFT ) to discover the unknown parameters in differential equations when only randomly distributed sparse data points are available. PIFT consists of three stages; (i) generating low fidelity data from prior knowledge under realistic settings, (ii) pre-training a single neural network with the generated low fidelity data, and (iii) fine-tuning the pre-trained model using physics-informed loss function. PIFT is evaluated on seven scientific problems including five ordinary differential equations and two partial differential equations. We also demonstrate the robustness and generalizability of PIFT to out-of-distribution tasks. PIFT exhibits high accuracy and robustness in discovering unknown parameters of differential equations from randomly distributed sparse data points.}, archive = {J_EAAI}, author = {Yong Jin Jeong and Taesup Moon}, doi = {10.1016/j.engappai.2025.112132}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112132}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Physics-informed fine-tuning for physics discovery from random and sparse data}, volume = {162}, year = {2025}, }
• (2025). Deep Q-learning with feature extraction and prioritized experience replay for edge node overload in edge computing. EAAI, 162, 112124. (www)

  @article{ , abstract = {Keeping track of the edge nodes’ status information is crucial, which is the condition of their available compute capacity as measured by their Age-of-Information. In Internet of Things-oriented edge computing systems, the computational and software-defined infrastructure resources are heterogeneous and subject to rapid change. Edge computing systems often face dynamic workloads and limited computational resources, leading to frequent node overload scenarios. These overloads degrade system responsiveness and service availability, especially in latency-sensitive applications. The scheduling of computation tasks would consider both current resource availability and predicted overload risks. An intelligent, adaptive method that learns optimal task allocation under resource constraints has the potential to boost the operational efficiency of Internet of Things-oriented edge computing systems. Addressing edge node overload is critical for the sustainable and scalable deployment of edge-based infrastructures. This research designs a resource-overloaded detection model specifically for diverse workloads in edge computing systems. The proposed Deep Reinforcement Learning model explores two significant challenges: the selection of pertinent feature sets from the workload resource utilization storage, and their classification of overload and detection of fatal failure of edge computing nodes. We propose a Deep-Q Network with a prioritized experience replay framework for edge node resource overload. The framework relies on feature learning using Linear Discriminant Analysis and Deep Q Network with a prioritized experience replay to efficiently indicate the overload status of edge nodes and reward the system with actions that enhance edge resources allocation. Deep-Q Network is well-suited for sequential decision-making in dynamic environments, while prioritized experience replay improves sample efficiency by focusing on updating on high-priority transitions with larger temporal-difference errors. Features are learned automatically from the edge node resource profiling data generated on a real edge-based container infrastructure. Linear Discriminant Analysis reduces the high-dimensional state space by emphasizing the most discriminative features for scheduling decisions. The infrastructure executes an intelligent inference of containerized applications considered as resource-intensive applications. When the feature extraction is added to the proposed deep reinforcement learning model, the overload classifier’s performance is improved. Comparing the model with selection to the one without it, the total accuracy and F1-score were improved by 1.3% and 1.4%, respectively.}, archive = {J_EAAI}, author = {Lionel Nkenyereye and Boon Giin Lee and Wan-Young Chung}, doi = {10.1016/j.engappai.2025.112124}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112124}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep Q-learning with feature extraction and prioritized experience replay for edge node overload in edge computing}, volume = {162}, year = {2025}, }
• (2025). STEval: A framework for evaluating spatio-temporal crime prediction models. EAAI, 162, 112123. (www)

  @article{ , abstract = {Spatio-Temporal predictive models are crucial for forecasting where and when crimes will occur, aiding public security organizations in resource allocation and crime prevention. Despite numerous literature proposals, a lack of standardized evaluation criteria hinders comparability and reliability. To address this, we propose STEval, a comprehensive and flexible evaluation framework for spatio-temporal predictive models. STEval consists of four modules: data preparation, spatial structure definition, model training, and model evaluation. The framework’s robustness was demonstrated using a 40-million-record crime dataset from Minas Gerais State (Brazil) across five experimental scenarios, including variations in temporal granularity, spatial resolution, and distribution shifts. Results demonstrate that no single model is universally superior, and the most appropriate model depends on the specific application context. For instance, Spatio-Temporal Kernel Density Estimation (STKDE) consistently achieved high Hit Rates (HR), often exceeding 0.98 in fine spatial resolutions (e.g., 100-square meters grid) for violent crimes, while Spatial-Temporal Autoregressive Integrated Moving Average (STARIMA) demonstrated strong performance in temporal granularity tests, reaching HRs of up to 0.65 for theft predictions. Conversely, Extra Tree Regressor, though exhibiting significantly lower HRs (e.g., as low as 0.02 in some spatial tests), consistently provided the fastest execution times, often under 5 s, contrasting with STKDE’s execution times that could extend to thousands of seconds in dense spatial grids. The framework’s detailed analysis provides important insights for informed model selection and optimization, revealing each model’s strengths and limitations, and providing a robust foundation for future advancements in spatio-temporal crime prediction research, leveraging Machine Learning (ML) techniques.}, archive = {J_EAAI}, author = {Gabriel Amarante and Matheus Pimenta and Yan Andrade and Matheus Senna and Rainer Menezes and Antônio Hot Faria and Marcelo Vilas-Boas and Frederico Martins de Paula Neto and João Paulo da Silva and Everton Renato de Sousa and Jamicel da Silva and Wagner Meira Jr. and George Teodoro and Leonardo Rocha and Renato Ferreira}, doi = {10.1016/j.engappai.2025.112123}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112123}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {STEval: A framework for evaluating spatio-temporal crime prediction models}, volume = {162}, year = {2025}, }
• (2025). A new design of arithmetic and logic unit for enhancing the security of future internet of things devices using quantum-dot technology. EAAI, 162, 112113. (www)

  @article{ , abstract = {The Internet of Things (IoT) is a network of interconnected devices that collect, monitor, analyze, and exchange data. This technology plays a crucial role in the smart city infrastructure by seamlessly interconnecting various nodes. The extensive application and recognition of IoT across multiple city domains, such as healthcare, transportation, energy, education, and agriculture, bring significant challenges, with security among the most pressing. Traditional hardware technologies like Complementary Metal Oxide Semiconductor (CMOS) and Very Large Scale Integration (VLSI) suffer from limitations such as high power consumption and insufficient scalability, which hinder secure and sustainable IoT deployment. Such limitations have prompted the need to seek other technologies that would serve the dual purpose of providing security as well as energy. Quantum-based technologies can become adequate candidates offering promising solutions to make IoT devices and sustainable systems more secured. Quantum-dot Cellular Automata (QCA) has been proposed as a nanotechnology with the potential of consuming ultra-low powers, less area, and high-speed operation. QCA enhances security through sustainable computing objectives by minimizing energy usage. To improve the future security and efficiency of IoT hardware, this paper suggests a QCA-based Arithmetic Logic Unit (ALU). This ALU can generate more than 12 logical and arithmetic operations. Designed together with the majority gates, XOR gates, multiplexers, and full adders, the ALU is simulated using the QCA-Designer 2.0.3. Simulated results indicate improvements in the number of cells and reduced occupied area relative to the earlier designs. These results indicate the potential of QCA technology in enabling secure, energy-efficient, and compact computing architecture applicable in the future IoT.}, archive = {J_EAAI}, author = {Maryam Zaker and Seyed Sajad Ahmadpour and Nima Jafari Navimipour and Muhammad Zohaib and Neeraj Kumar Misra and Sankit Kassa and Ahmad Habibizad Navin and Arash Heidari and Mehdi Hosseinzadeh and Omar I. Alsaleh}, doi = {10.1016/j.engappai.2025.112113}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112113}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A new design of arithmetic and logic unit for enhancing the security of future internet of things devices using quantum-dot technology}, volume = {162}, year = {2025}, }
• (2025). Prescribed-time convergence noise-tolerant zeroing neural network for multi-robot position management and coordination. EAAI, 162, 112072. (www)

  @article{ , abstract = {A zeroing neural network (ZNN) model with prescribed-time convergence and noise tolerance is constructed for the first time to address multi-robot position planning (MPP) tasks. Unlike traditional ZNN models, which converge within finite, fixed, or predefined times, this model ensures that system errors converge precisely within the predefined time T p . Furthermore, it significantly enhances noise tolerance while maintaining excellent convergence performance. Through rigorous mathematical analysis and detailed numerical simulations, the proposed model demonstrates exceptional convergence and robustness, even under noise disturbances. In a noise-free environment, the convergence time accuracy (CTA) of this model when solving the MPP task reach over 90%, which is approximately 30% higher than traditional predefined-time convergence neural networks.}, archive = {J_EAAI}, author = {Tinglei Wang and Yufei Wang and Cheng Hua and Xinwei Cao and Bolin Liao and Shuai Li}, doi = {10.1016/j.engappai.2025.112072}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112072}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Prescribed-time convergence noise-tolerant zeroing neural network for multi-robot position management and coordination}, volume = {162}, year = {2025}, }
• (2025). Inland waterway object detection in multi-environment: Dataset and approach. EAAI, 162, 111994. (www)

  @article{ , abstract = {Deep learning has advanced intelligent ship visual perception, but the scarcity of dedicated inland waterway vessels dataset limits system adaptability in complex environments. Narrow waterways, variable weather, and urban interference challenge the robustness of existing object detection systems. To address these issues, this paper constructs the Multi-environment Inland Waterway Vessels Dataset (MEIWVD), comprising 32,478 high-quality images from diverse Yangtze River basin scenarios, including sunny, rainy, foggy, and artificially lit conditions. The diversity and multi-scale characteristics of MEIWVD establish it as a rigorous benchmark for vessel detection. To leverage the characteristics of the MEIWVD, this paper proposes a scene-guided image enhancement module for multi-environment scenarios, which adaptively enhances water surface images based on environmental conditions to improve detector performance in complex scenarios. Additionally, a parameter-limited dilated convolution is introduced to enhance the representation of salient features of inland waterway vessels by leveraging their geometric characteristics. Finally, a multi-scale dilated residual fusion method is proposed to effectively integrate multi-scale features and improve the detection of multi-scale objects. Comprehensive statistical analysis and experiments on the MEIWVD demonstrate that it poses higher demands on object detection algorithms compared to existing water surface datasets, owing to its diverse and challenging scenarios. The proposed methods, including scene-guided image enhancement, parameter-limited dilated convolution, and multi-scale dilated residual fusion, advance research in multi-environment dataset, achieving a mean average precision over intersection over union thresholds from 0.5 to 0.95 of 81.6 % on the MEIWVD, outperforming state-of-the-art object detection algorithms.}, archive = {J_EAAI}, author = {Shanshan Wang and Haixiang Xu and Hui Feng and Xiaoqian Wang and Pei Song and Sijie Liu and Jianhua He}, doi = {10.1016/j.engappai.2025.111994}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111994}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Inland waterway object detection in multi-environment: Dataset and approach}, volume = {162}, year = {2025}, }
• (2025). GhostPointNet: A deep learning-based method for ghost point noise detection in four-dimensional (4D) millimeter-wave radar point clouds of underground mine. EAAI, 161, 112380. (www)

  @article{ , abstract = {The high dust concentration, multi-metal supports, and narrow winding tunnels in underground mines collectively lead to frequent ghost point noise in four-dimensional (4D) millimeter-wave radar point clouds, posing serious challenges for mining perception and localization. To address this, we propose a deep learning algorithm, named GhostPointNet, for 4D millimeter-wave radar ghost point detection in underground mining environments. From an artificial intelligence perspective, this model thoroughly considers the multi-modal features of 4D millimeter-wave radar and the environmental complexity of underground mines. It incorporates multi-parameterized spatial information inputs in both Cartesian and Spherical coordinates, coupled with “Double T-Net” adaptive alignment correction, while integrating non-spatial information such as radar power and Doppler data to achieve multi-modal representation and end-to-end discrimination between ghost points and real points. Experimental validation shows that GhostPointNet achieves excellent performance in underground mining scenarios with 92.45 % accuracy and 95.84 % F1-score, outperforming traditional filtering, clustering, and machine learning algorithms. From an engineering application perspective, GhostPointNet is specifically designed for ghost noise detection in underground mines. Even in complex scenarios such as mine tunnel intersections and turns, it preserves critical structural points. Its end-to-end neural network simplifies post-processing procedures, enhances operational efficiency, and provides stable and reliable perceptual support for subsequent tasks such as autonomous mine locomotive navigation and three-dimensional (3D) structure reconstruction. Experimental results demonstrate that this method surpasses baseline approaches in ghost point detection, real point preservation, and generalization capability, providing significant support for improving underground mining safety and efficiency.}, archive = {J_EAAI}, author = {Hu Liu and Zhenghua Zhang and Jing Yang and Jörg Benndorf and Xiaofei Wang and Jiaqi Dong and Zitao Lin and Guoliang Chen}, doi = {10.1016/j.engappai.2025.112380}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112380}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {GhostPointNet: A deep learning-based method for ghost point noise detection in four-dimensional (4D) millimeter-wave radar point clouds of underground mine}, volume = {161}, year = {2025}, }
• (2025). Hybrid ensemble learning for predicting peak deviatoric stress in soil-rock mixtures from triaxial test data. EAAI, 161, 112377. (www)

  @article{ , abstract = {Peak deviatoric stress ( q f ) is an important mechanical property index of soil-rock mixtures (SRM). However, determining it through experiments is a challenging and time-consuming task. This study presents a weighted averaging ensemble model for predicting the q f of SRM. 585 sets of consolidated drained triaxial test data for SRM were collected to construct the proposed models. Firstly, taking the extreme gradient boosting (XGBoost) as an example, the accuracy of the models established with different input parameter combinations was tested to determine the optimal inputs. Then, five artificial intelligence models were used to train the dataset, and the Bayesian optimization method was adopted for hyperparameter adjustment. Based on the prediction results, two base models (i.e. XGBoost and random forest (RF)) with good performance were selected from these five models, and a novel weighted averaging ensemble model was developed to predict the q f of SRM. The coefficient of determination (R 2 ), root mean squared error (RMSE) and mean absolute error (MAE) values of the ensemble model were 0.990, 220.0 kPa (kPa) and 118.4 kPa, respectively, for all datasets, indicating that this model has great potential in predicting the q f of SRM. In addition, the robustness, parameter patterns and the SHapley Additive exPlanations (SHAP) analysis of the ensemble model were also conducted in this study. The results of this study can be applied to rapid analysis of mechanical parameters for SRM in engineering, and are of great significance for geotechnical engineering design and disaster prevention.}, archive = {J_EAAI}, author = {Ruiliang Zhang and Xinhua Xue}, doi = {10.1016/j.engappai.2025.112377}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112377}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hybrid ensemble learning for predicting peak deviatoric stress in soil-rock mixtures from triaxial test data}, volume = {161}, year = {2025}, }
• (2025). Graph attention neural network for advancing medical imaging by enhancing segmentation and classification. EAAI, 161, 112372. (www)

  @article{ , abstract = {Brain tumor classification and segmentation using Magnetic Resonance Imaging (MRI) scans remain challenging due to limited spatial relationships and the inability of conventional models, such as Convolutional Neural Networks (CNNs), to effectively capture structural dependencies. To address these limitations, this study introduces GANN-Med (Graph Attention Neural Network for Medical Imaging), a novel framework designed to improve the accuracy and reliability of brain tumor detection, segmentation, and classification. The framework utilizes the publicly available Brain Tumor MRI dataset and applies preprocessing followed by wavelet transform to extract both spatial and frequency features, generating wavelet-based node embeddings that represent fine textures and overall tumor structure. These features are modeled as a graph and processed through a U-Net (U-shaped Convolutional Network) architecture integrated with Graph Attention Networks (GATs), enabling adaptive attention to be directed to critical tumor regions. The segmented tumor regions are then classified using GraphSAGE (Graph Sample and Aggregation) with a pooling aggregator, which supports inductive neighborhood learning while preserving topological consistency. Unlike existing models such as mResU-Net (Modified Residual U-Net) and LG-GNN (Local-Global Graph Neural Network), GANN-Med uniquely combines wavelet-based multi-resolution graph embeddings with attention mechanisms in a unified architecture. Experimental results demonstrate a significant performance boost, including an 18.47 % increase in Dice Similarity Coefficient (DSC), notable improvements in Jaccard Index and sensitivity, a 21.04 % reduction in false positives, and an overall classification accuracy of 93.2 %, with a balanced accuracy of 91.8 % across glioma, meningioma, pituitary, and no tumor classes. Additionally, the model achieves minimal training and validation losses of 43.2 % and 41.6 %, respectively, highlighting its potential to minimize misdiagnosis and contribute to more accurate, early-stage clinical decision-making in neuro-oncology.}, archive = {J_EAAI}, author = {Meshari D. Alanazi and Khaled Kaaniche and Mohammed Albekairi and Turki M. Alanazi and Munid Alanazi and Ghulam Abbas}, doi = {10.1016/j.engappai.2025.112372}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112372}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Graph attention neural network for advancing medical imaging by enhancing segmentation and classification}, volume = {161}, year = {2025}, }
• (2025). Fuzzy scheduling in distributed heterogeneous printed circuit board assembly lines: Feedback assisted neighborhood-based search coupling with rapid evaluations. EAAI, 161, 112353. (www)

  @article{ , abstract = {Printed Circuit Board (PCB) production involves multiple heterogeneous and highly automated assembly lines running concurrently. These assembly lines involve complex processes with uncertain and fuzzy production times, presenting significant challenges for scheduling. Moreover, PCB production is notably affected by carryover sequence-dependent setup times (CSDST) and the need for product grouping, further complicating the scheduling process. To address these challenges, this paper formulates the scheduling problem as a fuzzy distributed heterogeneous flowshop group scheduling problem with carryover sequence-dependent setup times (FDHFGSP/CSDST). A novel and effective feedback-assisted population-based neighborhood search (FPBNS) algorithm, coupling with rapid evaluation of neighboring solutions, is proposed to solve this problem. The feedback strategy operates on two main levels: evaluating population quality and adjusting the search process. Additionally, an integrated heuristic method is employed to generate a high-quality and diverse initial population. To further enhance solution quality, a collaborative operation is designed to improve interaction between solutions. This comprehensive algorithm significantly enhances its adaptability in PCB manufacturing, providing an efficient solution to the complex scheduling challenges of real-world production environments. Comprehensive and systematic experiments were conducted to evaluate the effectiveness of the proposed algorithm and related methods, demonstrating its superior performance in solving the complex scheduling problem in PCB manufacturing.}, archive = {J_EAAI}, author = {Zhen-duo Han and Biao Zhang and Chao Lu and Lei-lei Meng and Wen-qiang Zou}, doi = {10.1016/j.engappai.2025.112353}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112353}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fuzzy scheduling in distributed heterogeneous printed circuit board assembly lines: Feedback assisted neighborhood-based search coupling with rapid evaluations}, volume = {161}, year = {2025}, }
• (2025). A high-precision acupoint recognition and localization model for acupuncture robot end-effectors. EAAI, 161, 112348. (www)

  @article{ , abstract = {Accurate acupoint recognition is fundamental to therapeutic efficacy of acupuncture robots, presenting a crucial challenge in medical robotics: high-precision, real-time localization of sparse features on deformable skin surfaces. To address this, we propose the Acupoint Recognition Network (ACU-Net) for reliable acupoint identification to guide robotic acupuncture. From an artificial intelligence (AI) perspective, the core contribution of ACU-Net lies in its novel modules. The Convolutional Dilatation with Residual and Split (CDRS) module uses dilated convolution to expand the receptive field and overcomes the sparsity of acupoint features by capturing both anatomical contexts and subtle local gradients. The Cross-Scale Feature Fusion (CSFF) architecture integrates spatial and semantic information more effectively than standard Feature Pyramid Network (FPN). Formulated as a keypoint regression task, the network is jointly optimized using a weighted Object Keypoint Similarity (OKS)-based loss function. Evaluated on a physician-annotated, large-scale chest acupoint dataset, ACU-Net has achieved a mean Average Precision (AP 50-95 (B)) of 94.1 % for acupoint bounding boxes and 99.5 % for acupoint points (AP 50-95 (P)), outperforming mainstream Convolutional Neural Networks (CNNs) and Transformer models. Its average prediction error (0.282 cm) on volunteers represents a 30.9 % reduction compared to the second-best model of Real-Time Multi-person Pose Estimation (RTMPose). Robustness test on an independent forearm dataset confirmed its precision. Attention visualization showed strong focus on acupoint regions. Testing on chest acupoint under adverse imaging conditions of low illumination and information loss demonstrated its reliability with zero missed detection. This work significantly advances the development of acupuncture robotics for intelligent diagnosis and treatment.}, archive = {J_EAAI}, author = {Ailing Tan and Hao Mu and Wei Ma and Yu'e Lv and Haoyu Wang and Yong Zhao}, doi = {10.1016/j.engappai.2025.112348}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112348}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A high-precision acupoint recognition and localization model for acupuncture robot end-effectors}, volume = {161}, year = {2025}, }
• (2025). Eviformer: An uncertainty fault diagnosis framework guided by evidential deep learning. EAAI, 161, 112328. (www)

  @article{ , abstract = {Unpredictable signals are commonly encountered during equipment operation, and existing deep learning-based fault diagnosis methods often fail to accurately evaluate the uncertainty of diagnostic results, limiting the model's capacity to respond to unexpected signals. To address this challenge, this study introduces a modified Transformer model based on deep evidential learning—Eviformer. The proposed model first employs a Swin-Transformer (ST)-based distribution projector, which preserves the advantages of ST in extracting features from vibration signals and simultaneously projects these signals directly into a Dirichlet distribution with second-order probabilities. Furthermore, by incorporating novel evidence correction terms and a constraint factor to reconstruct the evidence constraint loss, more precise uncertainty quantification in diagnostic predictions is achieved. Using a gear-bearing vibration dataset, comparative experiments were conducted across various scenarios, including out-of-distribution gear faults, faults in unmonitored components, noise interference, and variable speed conditions. The results demonstrate that the proposed method can promptly issue uncertainty-based warnings when encountering vibration signals that significantly differ from the training set distribution, thereby offering essential support for maintenance decisions.}, archive = {J_EAAI}, author = {Jingjie Luo and Fucai Li and Xiaolei Xu and Wenqiang Zhao and Dongqing Zhang}, doi = {10.1016/j.engappai.2025.112328}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112328}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Eviformer: An uncertainty fault diagnosis framework guided by evidential deep learning}, volume = {161}, year = {2025}, }
• (2025). Multi-scale spatiotemporal feature fusion for super-resolution video reconstruction in dynamic scenes. EAAI, 161, 112327. (www)

  @article{ , abstract = {High-quality geospatial video reconstruction is vital for applications such as urban surveillance and disaster assessment. However, existing methods often suffer from motion blur and visual artifacts due to hardware limitations and dynamic environments with rapid motion, frequent background changes, and lighting variations. To address these challenges, we propose a novel video super-resolution method called Multi-Scale Spatiotemporal Feature Fusion for Video Super-Resolution. This approach leverages artificial intelligence techniques to extract rich spatiotemporal features using a multi-scale deformable convolutional pyramid with parallel branches. It integrates contextual residual learning and optical flow estimation to capture spatial details and compensate for missing temporal information. A hybrid spatiotemporal attention mechanism adaptively fuses key features across frames through dynamic weighting, suppressing background interference and enhancing target-region representation. The framework includes an explicit motion compensation module for precise frame alignment, followed by upsampling to generate high-resolution frames with improved perceptual quality. A multi-task joint loss function optimizes structural fidelity, motion consistency, and visual realism. Experimental results on the University of Alberta Detection and Tracking dataset and Video for Super-Resolution benchmark show the method achieves a peak signal-to-noise ratio of 30.62 dB and a structural similarity index of 0.92, surpassing state-of-the-art methods by up to 4.28 dB and improving perceptual metrics such as Learned Perceptual Image Patch Similarity and Flip Loss in Perception. Ablation studies confirm each module's contribution. The model runs efficiently at 24.8 ms per frame, demonstrating potential for real-time applications. This approach offers a robust and efficient solution for high-resolution geospatial video reconstruction.}, archive = {J_EAAI}, author = {mengqin Yang and yanhui Wang and yang Yang and chunhua Peng}, doi = {10.1016/j.engappai.2025.112327}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112327}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-scale spatiotemporal feature fusion for super-resolution video reconstruction in dynamic scenes}, volume = {161}, year = {2025}, }
• (2025). Explainable framework based on integrated sensor fusion and deep learning to estimate stability in high-speed micromilling of thin-walled titanium alloy. EAAI, 161, 112326. (www)

  @article{ , abstract = {A novel deep learning (DL) framework based on multi-sensor fusion based on vibration and audio signals with transfer learning has been developed and validated for chatter onset prediction in micromilling of thin-walled titanium (TC4) alloys. The DL framework has been trained and validated with individual sensor data, labelled using machined surface micrographs and cutting tool condition. A total of 879,840 spectrograms has been generated for training via signal segmentation and augmentation methods that yielded a validation accuracy of 89 % during training phase. The trained DL model has been tested on the unlabelled fused signal data to make the DL model adapt to the unseen patterns in the signals recorded during real-time machining. The proposed DL framework based on fused signals outperformed individual sensor-based detection methods by demonstrating an increment of 33 % in accuracy, with a mean absolute error of 0.027416 and R 2 of 0.987. Ablation studies confirm that fusing signals enhances the DL model's performance, yielding a 21 % and 29 % improvement in precision and recall, respectively. The optimal fusion strategy reduced chatter misclassification, with one instance of misclassification. The proposed framework presents a robust, explainable DL approach that can be readily used for adaptive control of cutting conditions in different manufacturing processes. The industry readiness of the DL model has been demonstrated by testing the pretrained model on turning process. The DL model shows 95.31 % prediction accuracy with only three misclassifications out of sixty-four test conditions.}, archive = {J_EAAI}, author = {Sethurao Gururaja and Kundan K. Singh}, doi = {10.1016/j.engappai.2025.112326}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112326}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Explainable framework based on integrated sensor fusion and deep learning to estimate stability in high-speed micromilling of thin-walled titanium alloy}, volume = {161}, year = {2025}, }
• (2025). Liquid metal microfluidic cooling system for high-efficiency thermal management via learning-based genetic algorithm. EAAI, 161, 112324. (www)

  @article{ , abstract = {High heat flux density is a critical factor that limits the performance and reliability of miniaturized, high-power microelectronic systems. This study proposes a liquid metal (LM)-based microfluidic cooling system optimized through a data-driven computational framework based on an enhanced Genetic Algorithm (LC-GA), aiming to deliver an efficient thermal management solution for high-density integrated systems. By integrating LM near-junction cooling with microchannel heat dissipation in a silicon substrate, we developed a heterogeneous three-dimensional interconnect cooling architecture capable of optimizing thermal performance through algorithm-guided parameter tuning. To validate the proposed method, four distinct microchannel configurations were designed, fabricated, and experimentally tested. LM was introduced into the channels to conduct both experimental cooling tests and thermal performance simulations on a simulated heat source. The results demonstrate that this LM-based microfluidic cooling system, optimized through computational parameter determination, can effectively dissipate heat from chips with power consumption up to 800 W while maintaining stable thermal performance. Additionally, a response surface methodology combined with enhanced LC-GA was utilized for multi-factor sensitivity analysis and multi-objective optimization, enabling automatic determination of optimal design and operating parameters to balance thermal resistance and pressure drop. The optimized configuration reduced the maximum chip temperature to approximately 357.54 K, lowered the system pressure requirement, and improved the Performance Evaluation Criterion (PEC) to 2.327. This work provides a data-driven optimization approach that supports the development of high-performance integrated microsystems through algorithm-assisted thermal design.}, archive = {J_EAAI}, author = {Yucheng Wang and Antong Bi and Kaiyu Chen and Shenxin Yu and Wanping Gao and Wenyi Zhang and Yuwan Wu and Zhiqiang Li and Shaoxi Wang}, doi = {10.1016/j.engappai.2025.112324}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112324}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Liquid metal microfluidic cooling system for high-efficiency thermal management via learning-based genetic algorithm}, volume = {161}, year = {2025}, }
• (2025). HoloFluoNet: Live cell imaging intelligence based on fused holography and fluorescence for virtual staining, cell segmentation, classification, and viability analysis. EAAI, 161, 112323. (www)

  @article{ , abstract = {Fluorescence microscopy enables detailed visualization of subcellular structures but is limited by phototoxicity, photobleaching, and labor-intensive labeling. In contrast, digital holographic microscopy (DHM) offers label-free, quantitative phase imaging but lacks biochemical specificity. To integrate the strengths of both modalities, we propose HoloFluoNet, a deep learning-based framework that generates virtual fluorescence information from phase images acquired by DHM. Using a dual-mode imaging system, we simultaneously captured phase and fluorescence images of live cancer cells. The fluorescence data were used to construct biologically grounded supervision signals, including nuclei masks and multi-class cell viability labels. From a single-phase image, HoloFluoNet predicts a virtual nuclei mask, a distance map for boundary refinement, and a viability mask. The architecture incorporates multi-scale feature extractor, and attention mechanisms, optimized with novel inclusion and exclusion loss functions. Post-processing using the watershed algorithm ensures accurate segmentation of overlapping cells. The final registered image provides label-free virtual staining for nuclei localization, viability assessment, and class-specific morphological profiling. HoloFluoNet achieved a Dice score of 0.8685 and an Aggregated Jaccard Index (AJI) of 0.7883, outperforming conventional deep learning models such as U-Net, Fully Convolutional Network (FCN), Pyramid Scene Parsing Network (PSPNet), and DeepLab v3+. These improvements were statistically validated using one-way analysis of variance (p < 0.01). Ablation experiments confirmed the complementary roles of architectural modules and novel loss functions, while robustness tests under noisy and low-quality conditions revealed high stability to low contrast and moderate noise. With an inference speed of 0.123 s per image, the model enables real-time cellular analysis. By bridging structural and molecular imaging, HoloFluoNet provides an efficient, label-free alternative to fluorescence microscopy, with promising applications in artificial intelligence (AI)-assisted drug screening, cancer research, and live-cell monitoring.}, archive = {J_EAAI}, author = {Seonghwan Park and Jaeseong Lee and Jaewoo Park and Inkyu Moon}, doi = {10.1016/j.engappai.2025.112323}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112323}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {HoloFluoNet: Live cell imaging intelligence based on fused holography and fluorescence for virtual staining, cell segmentation, classification, and viability analysis}, volume = {161}, year = {2025}, }
• (2025). Method for ensuring quasi-maximum stability of a system with interval deviations by a robust controller parametric synthesis. EAAI, 161, 112320. (www)

  @article{ , abstract = {It is known that ensuring reliable stability of automatic control systems with interval and stochastic parameter deviations is critical for aviation and other critical areas. Therefore, this article proposes a method for synthesizing Proportional-Integral-Derivative controller parameters based on the coefficient approach to stability analysis (through the characteristic polynomial neighboring coefficients ratios) with an extension to interval uncertainties and the stability quasi-maximum degree integration into the multi-criteria optimization Non-Dominated Sorting Genetic Algorithm II. Scientific contribution and novelty lie in combining analytical sufficient stability conditions (including verification by Kharitonov polynomials and dominant polynomial analysis) with multi-objective optimization taking into account parameter errors and energy costs, which allows obtaining stable and economical Proportional-Integral-Derivative settings for discrete two-channel systems. The results are confirmed by numerical experiments on a helicopter turboshaft engine two-channel control system model with real on-board data. It is shown that the optimal parameter vector provides overshoot of ≈1.43 % and transient process time of ≈1.12 s and maintains stability with local changes in coefficients (±3 … 7 %) and in Monte-Carlo tests. At the same time, a comparative analysis with traditional and neural network controllers showed the proposed approach's advantage in key metrics.}, archive = {J_EAAI}, author = {Denys Baranovskyi and Serhii Vladov and Valerii Sokurenko and Oleksandr Muzychuk and Victoria Vysotska and Maryna Bulakh}, doi = {10.1016/j.engappai.2025.112320}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112320}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Method for ensuring quasi-maximum stability of a system with interval deviations by a robust controller parametric synthesis}, volume = {161}, year = {2025}, }
• (2025). A robust deep learning model selection with data augmentation for automatic detection of tessellated fundus images and explainable artificial intelligence based interpretation. EAAI, 161, 112316. (www)

  @article{ , abstract = {A robust deep learning system for automatically classifying retinal fundus images into two classes—normal and tessellated—is presented in this study. Visual Geometry Group – 16 is used as the base model, taking advantage of transfer learning to develop an efficient framework for fundus image classification. The approach uses nine different model architectures and makes use of a dataset of 352 fundus images that was increased to 4865 samples using sophisticated data augmentation techniques. Of these, Model_8 performed the best, achieving a loss of 0.129 % and an impressive accuracy of 99.39 %. The suggested approach ensures higher performance and dependability by combining rigorous data augmentation, efficient preprocessing, and model fine-tuning techniques. Furthermore, Explainable Artificial Intelligence was used to improve the interpretability of the model and visualize important aspects such as features or imposed pathologies in fundus images more clearly. The study offers promising support to ophthalmologists by offering precise automated diagnoses for the early identification and treatment of retinal disorders.}, archive = {J_EAAI}, author = {Kachi Anvesh and Shanmugasundaram Hariharan and Bharati M. Reshmi and Qiang Xu and Joan Lu and Vinay Kukreja and Murugaperumal Krishnamoorthy}, doi = {10.1016/j.engappai.2025.112316}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112316}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A robust deep learning model selection with data augmentation for automatic detection of tessellated fundus images and explainable artificial intelligence based interpretation}, volume = {161}, year = {2025}, }
• (2025). Camera-invariance correlation learning and inter-domain-specific distinct representation for person re-identification. EAAI, 161, 112314. (www)

  @article{ , abstract = {Unsupervised domain adaptation (UDA) in person re-identification (ReID) experiences performance degradation due to two primary factors: (1) the information changes across cameras among intra-domain identity samples within the source and target domains; and (2) the inter-domain gap when applied to an unexplored image domain, which makes it a challenging task. The majority of current methodologies predominantly address the inter-domain gap while overlooking the impact of the intra-domain issue. To overcome this limitation, this paper proposes an effective invariance representation ReID (IR-ReID) network that encompasses camera-invariance correlation learning (CICL) and Transformer-based, inter-domain-specific discriminative representation (IDSR), supported by curriculum learning, with the aim of concurrently addressing both of these pivotal issues. Specifically, a CICL module is designed to explore intra-domain distinct features of an individual captured by multiple cameras across the source and target domains. Secondly, an IDSR module is proposed to address the inter-domain gap issue. Thirdly, a curriculum learning-driven strategy is introduced for sample training, systematically guiding our network in learning samples progressively, ranging from easy to difficult instances. Meanwhile, a cluster-based constraint loss is presented to further optimize the clustering results. Finally, extensive experiments are conducted on three well-known datasets, which demonstrate the effectiveness and superiority of our proposed approach in comparison to the prevailing state-of-the-art methodologies.}, archive = {J_EAAI}, author = {Shangdong Zhu and Yunzhou Zhang and Peng Duan}, doi = {10.1016/j.engappai.2025.112314}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112314}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Camera-invariance correlation learning and inter-domain-specific distinct representation for person re-identification}, volume = {161}, year = {2025}, }
• (2025). Applications of machine vision technology for conveyor belt deviation detection: A review and roadmap. EAAI, 161, 112312. (www)

  @article{ , abstract = {Conveyor belt deviation is a frequent challenge in product transportation filed, and failure to promptly detect and rectify this anomaly not only significantly reduces transport efficiency but also poses a risk of serious safety accidents, leading to enormous economic losses. Traditional contact-based deviation detection technologies, with their inherent limitations of high costs and complicated maintenance, have struggled to meet the practical demands of long-distance conveyor belt inspection. In this context, non-contact machine vision technology has emerged as a prominent solution in the field of conveyor belt deviation detection, thanks to its notable advantages of a simple hardware structure and round-the-clock operational capability. Recently, with the rapid development of artificial intelligence theories, this research field has accumulated a series of effective solutions that have been proven through machine vision practical applications. This paper delves into the technical principles of the existing solutions, systematically summarizes them and objectively evaluates their strengths and weaknesses in practical applications. Based on this foundation, this paper also provides an insight on the future development trends of intelligent monitoring for conveyor belt deviation, aiming to offer valuable reference and guidance to technicians in related fields.}, archive = {J_EAAI}, author = {Jiaming Han and Ting Fang and Wensheng Liu and Chenxiao Zhang and Molin Zhu and Jibin Xu and Jie Ji and Xianhua He and Zhang Wang and Min Tang and Chong Dong and Long Ma and Xinlong Yang}, doi = {10.1016/j.engappai.2025.112312}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112312}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Applications of machine vision technology for conveyor belt deviation detection: A review and roadmap}, volume = {161}, year = {2025}, }
• (2025). Synthetic aperture radar target recognition with limited training data based on frequency-domain-assisted dual-stream attention hierarchical deformable convolutional networks. EAAI, 161, 112309. (www)

  @article{ , abstract = {In the field of synthetic aperture radar (SAR) automatic target recognition (ATR), the recognition performance of deep learning methods is often constrained by sample scarcity. To address this issue, this paper proposes a frequency-domain-assisted dual-stream attention hierarchical deformable convolutional network (DAHDF-Net). A frequency-domain-assisted dataset construction method is designed to extend the representational diversity of the sample feature space. Then, a parallel input architecture of frequency-space dual-stream features is constructed, and the adaptive dynamic fusion of frequency-domain stream features with spatial-domain stream features is realized by using a channel attention module in the frequency-domain stream. To enhance the expressiveness of the network on the basis of modulated deformation convolution, a hierarchical residual (H-residual) module is designed. Moreover, a triple hybrid loss is proposed to address the issues of high deformation, which makes classification difficult, and the fuzzy decision boundaries between categories. Experiments were conducted on the moving and stationary target acquisition and recognition (MSTAR) dataset and the FUSAR-Ship dataset. Among them, DAHDF-Net achieves recognition accuracies of 99.75 % and 97.48 % on all training sets and 10-way 25-shot under standard operating condition (SOC) of MSTAR, which demonstrated superior recognition accuracy compared to current state-of-the-art methods.}, archive = {J_EAAI}, author = {Shanliang Yuan and Hongyuan Gao and Xiaoyuan Gu and Jingya Ma}, doi = {10.1016/j.engappai.2025.112309}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112309}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Synthetic aperture radar target recognition with limited training data based on frequency-domain-assisted dual-stream attention hierarchical deformable convolutional networks}, volume = {161}, year = {2025}, }
• (2025). Multi-scale distance similarity entropy: A novel complexity measurement for gearbox fault diagnosis. EAAI, 161, 112308. (www)

  @article{ , abstract = {Entropy-based methods have gained increasing attention in fault diagnosis due to their capability in quantifying the intrinsic complexity and dynamic behavior of time series signals. Vibration signals from gearboxes are typically nonlinear, nonstationary, and transient, often spanning multiple time scales, thus requiring robust multi-scale analysis for accurate condition assessment. However, existing multi-scale entropy methods, prone to missing local variations and vulnerable to noise, often yield incomplete features and low diagnostic accuracy with high computational costs. To overcome these challenges, this paper proposes a novel complexity measurement method termed multi-scale distance similarity entropy, which integrates distance similarity entropy with a multi-scale coarse-graining process. By computing element-wise distances and mapping them through a Gaussian kernel function, multi-scale distance similarity entropy captures subtle nonlinear variations in vibration signals while effectively suppressing noise interference. It estimates the similarity distribution between adjacent vectors, enabling sensitive tracking of dynamic complexity changes in mechanical vibration data. Experimental validation on two gearbox datasets encompassing both single and compound faults demonstrates that multi-scale distance similarity entropy achieves superior classification accuracies of 98.63 % and 97.14 %, respectively. The method also shows strong robustness under random impulse noise and low signal-to-noise ratio conditions, along with high computational efficiency, particularly in small-sample scenarios. These characteristics make it highly suitable for online real-time fault diagnosis in complex and noise contaminated industrial environments involving diverse and compound fault patterns.}, archive = {J_EAAI}, author = {Tao Wang and Shin Yee Khoo and Zhi Chao Ong and Pei Yi Siow and Teng Wang}, doi = {10.1016/j.engappai.2025.112308}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112308}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-scale distance similarity entropy: A novel complexity measurement for gearbox fault diagnosis}, volume = {161}, year = {2025}, }
• (2025). JSSM-DTA: Joint sequence-structure modeling with multi-scale transformers for explainable drug-target affinity prediction. EAAI, 161, 112303. (www)

  @article{ , abstract = {Drug-Target Affinity (DTA) prediction is a pivotal task in drug discovery, aiming to quantify the strength of interactions between drug molecules and their biological targets, such as proteins. Despite significant advancements in deep learning-based methods, several challenges persist. Scale mismatches obstruct functional group-specific representations, leading to suboptimal feature extraction and loss of critical interaction patterns. Additionally, the inability to fully exploit molecular graph structures weakens hierarchical dependency modeling, while ineffective feature fusion hampers the integration of heterogeneous molecular representations, restricting the capture of intricate drug-target interaction dynamics. To address these challenges, this work introduces JSSM-DTA, a novel Joint Sequence-Structure Modeling framework that integrates Multi-Scale Transformers with intermodal feature attention to enhance DTA prediction. JSSM-DTA formulates a unified representation space through the Adaptive Convolutional Transformer (ACT) and the Multi-Scale Diffusion Transformer (MSDT), where ACT refines sequential feature extraction by capturing intricate local dependencies and long-range contextual relationships, while MSDT preserves hierarchical learning, optimizing both global topological organization and fine-grained interactions. Additionally, the Factorized Inter-layer Interaction Module constructs joint embeddings by seamlessly integrating heterogeneous representations, enabling robust feature fusion. The proposed model was evaluated on Davis, KIBA, Metz and BindingDB datasets, demonstrating superior predictive accuracy over state-of-the-art methods. Explainability analysis reveals key molecular substructures and binding interactions, enhancing transparency in decision-making and establishing JSSM-DTA as a benchmark for both accuracy and interpretability in DTA prediction.}, archive = {J_EAAI}, author = {Uma E. and Mala T.}, doi = {10.1016/j.engappai.2025.112303}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112303}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {JSSM-DTA: Joint sequence-structure modeling with multi-scale transformers for explainable drug-target affinity prediction}, volume = {161}, year = {2025}, }
• (2025). Cyclic translations between pathomics and genomics improve automatic cancer diagnosis from whole slide images. EAAI, 161, 112302. (www)

  @article{ , abstract = {Incorporating genomic characterization into histopathological image modeling brings substantial value to enhancing diagnostic accuracy and supporting the development of targeted and effective treatment strategies. However, prevailing multi-modal integration methods often assume the availability of both pathomics and genomics data in both training and testing phases, overlooking the challenge of data absence due to prohibitive costs. In this paper, we propose a multi-modal cyclic feature generation network (MCFGN) that facilitates cyclic translations between pathomics and genomics to acquire a unified representation of multi-modal data. This approach enables the use of pathological images alone as input to generate joint representations during the testing phase. First, we utilize a general-purpose, self-supervised vision encoder to embed histological image patches as distinctive visual tokens. Next, we hierarchically aggregate patch-level tokens to region-level and slide-level, generating improved whole slide image (WSI) representations. We build self-supervised Masked Autoencoders (MAE) to initialize the hierarchical aggregator. Finally, to incorporate genomic characterization into the learning process, we develop a novel cross-modal cyclic feature generation module to create an intermediate joint representation of pathological and genetic features for patient diagnosis. Evaluations have been conducted on two public datasets from The Cancer Genome Atlas Breast Invasive Carcinoma (TCGA-BRCA) and Non-Small Cell Lung Cancer (TCGA-NSCLC) for various diagnostic tasks, including cancer subtyping and biomarker status prediction. Experiments indicate that our MCFGN model improves predictive performance in cancer diagnosis using histological slides, yielding an 8.7% improvement in area under the curve (AUC) for the cancer subtyping task and a 14.1% gain for biomarker prediction.}, archive = {J_EAAI}, author = {Xinyu Hao and Hongming Xu and Xiaofeng Wang and Tong Wang and Timo Hamalainen and Fengyu Cong}, doi = {10.1016/j.engappai.2025.112302}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112302}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Cyclic translations between pathomics and genomics improve automatic cancer diagnosis from whole slide images}, volume = {161}, year = {2025}, }
• (2025). A multimodal industrial anomaly detection method based on mask training and teacher–student joint memory. EAAI, 161, 112299. (www)

  @article{ , abstract = {In recent years, the teacher–student framework has been applied to both single-modality detection and multimodal detection, which realizes anomaly detection based on the feature difference between the teacher model and the student model. However, current multimodal teacher–student models use the same teacher model to extract two-dimensional (2D) image and three-dimensional (3D) point cloud features. The point cloud features extracted by the teacher model pre-trained on images are not the optimal feature representation. To further improve the performance of the teacher–student framework on the multimodal anomaly detection task, this paper proposes M ultimodal T eacher- S tudent J oint M emory ( MTSJM ). MTSJM constructs a teacher–student joint memory bank for each modality, the feature distance between the test sample and the memory bank is used as the anomaly indicator. This distance reflects the feature differences between the test sample and the normal sample at multiple levels, including the teacher–teacher, teacher–student, and student–student levels. Then, this paper proposes a mask-based student model training method. While ensuring that the student learns the feature of normal regions, mask training increases the feature difference of non-normal regions between the student and the teacher. On the MVTec 3D Anomaly Detection (MVTec 3D-AD) dataset, the proposed MTSJM achieves effective anomaly detection performance, reaching 95.7% mean Image-level Area Under the Receiver Operator Curve (I-AUROC) and 97.2% mean Area Under the Per-Region Overlap (AUPRO). In addition, MTSJM achieves 99.3% I-AUROC and 99.6% Pixel-level AUROC (P-AUROC) on a real-world vehicle stamping part task, which further illustrates the applicability of MTSJM on the multimodal anomaly detection task.}, archive = {J_EAAI}, author = {Yi Liu and Changsheng Zhang and Xingjun Dong and Yufei Yang}, doi = {10.1016/j.engappai.2025.112299}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112299}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multimodal industrial anomaly detection method based on mask training and teacher–student joint memory}, volume = {161}, year = {2025}, }
• (2025). Fuse MetaFormer with convolutional neural networks for three-dimensional model classification. EAAI, 161, 112298. (www)

  @article{ , abstract = {Multimedia technology is widely applied to artificial intelligence and it is key to the performance of artificial intelligence systems, such as computer-aided design system, virtual reality system, the augmented reality system, medical image processing system, and game development system. With the development of multimedia technology, the number of three-dimensional (3D) models in network or database is becoming larger and larger. It is important to classify 3D models. In order to improve accuracy of three-dimensional model classification, a method of 3D model classification fusing MetaFormer with CNN (Convolutional Neural Networks) is proposed. 3D model is projected into two-dimensional (2D) views through the fixed-view projection, and representative views are selected by the clustering algorithm. Points sampled randomly from representative view are used to calculate its shape descriptors. View feature is extracted from representative view by MetaFormer. Shape feature is extracted from shape descriptors of representative view by Convolutional Neural Networks. View feature and shape feature of representative view are fused. At the same time, majority voting algorithm is used to determine category of 3D model based on the fusion of view features and shape features. Experiments are conducted on ModelNet10 dataset. Experimental results show that the proposed method achieves better results than others.}, archive = {J_EAAI}, author = {Xueyao Gao and Haomin Wu and Chunxiang Zhang and Yongzeng Xue}, doi = {10.1016/j.engappai.2025.112298}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112298}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fuse MetaFormer with convolutional neural networks for three-dimensional model classification}, volume = {161}, year = {2025}, }
• (2025). Intelligent design of broadband metasurface based on spectrum prediction neural network and transition simulated annealing algorithm. EAAI, 161, 112297. (www)

  @article{ , abstract = {Recently, metasurfaces have demonstrated considerable potential for application across various fields, including wireless communication, Internet of Things, holographic imaging, and radar systems. With the rapid development of artificial intelligence and its application in other disciplines, metasurface design based on artificial intelligence techniques, particularly deep learning, has attracted widespread attention. In this paper, we propose a novel scheme that integrates deep neural networks with an enhanced simulated annealing algorithm to achieve intelligent design of broadband metasurfaces. Firstly, we construct and train a neural network capable of accurately predicting the amplitude and phase responses of an isotropic metasurface structure, achieving a Mean Absolute Error of 1.36° on the test set. Subsequently, we develop a novel simulated annealing algorithm that is integrated with the prediction neural network to form an intelligent optimization framework. This approach can effectively search for the solution space to accurately achieve arbitrary target phase response, with a very low Mean Squared Error calculated on the test set. Furthermore, we validate the effectiveness of the proposed model and algorithm through a practical application that demonstrates broadband dispersion characteristics within the frequency range of 9–11 Gigahertz (GHz). Additionally, the prediction model can employ transfer learning strategy to adapt to various operating frequency bands, significantly enhancing its generalization capabilities. The proposed intelligent design methodology introduces a novel pathway for rapidly designing metasurface in complex application scenarios, serving as a valuable reference for the engineering design of devices in electronics, optics, and wireless communication.}, archive = {J_EAAI}, author = {Yin Zhang and Yuxin Dai and Yiyu Fan and Jun Yu}, doi = {10.1016/j.engappai.2025.112297}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112297}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Intelligent design of broadband metasurface based on spectrum prediction neural network and transition simulated annealing algorithm}, volume = {161}, year = {2025}, }
• (2025). Industrial fabric defect-generative adversarial network (IFD-GAN): High-fidelity fabric cross-scale defect samples synthesis method for enhancing automated recognition performance. EAAI, 161, 112296. (www)

  @article{ , abstract = {—High-quality defect samples and datasets are the cornerstone for enhancing the performance of deep learning detection algorithms for fabric defects. Constraints, such as the complexity of fabric textures, the diversity and sporadic occurrence of defects, the insufficient scale, and the significant class imbalance of datasets, impede improvements in detection accuracy and generalization of detection models, thereby limiting industrial application. To address the issues above, a two-stage generative model for fabric defect images called Industrial Fabric Defect-Generative Adversarial Network (IFD-GAN), based on contrastive learning mutual information mechanism, was introduced. The Pixel-level Defect-Background Stripper (DBS) was devised for precise localization, and the Longitude-Latitude Self-Attention Mechanism (LLSA) was proposed to efficiently focus on defect foreground details, facilitating decoupling and efficient extraction of defect features. The combination of the two, along with the structural similarity loss function, collectively regulates the coordination and consistency between defect and background textures. An industrial-grade fabric defect dataset was collected for IFD-GAN training and generative testing. The model's effectiveness in generating defect images was evaluated based on dimensions such as fidelity and diversity. IFD-GAN is used to enhance this dataset, and the system compared the performance of advanced object detection models trained on the datasets before and after the enhancement. Extensive experimental results show that IFD-GAN can accomplish high-fidelity generation of high-resolution cross-scale fabric defect images in an unsupervised environment, significantly contributing to the creation of more balanced and diverse large-scale industrial fabric defect datasets and the enhancement of defect detection networks in precision, robustness, and generalizability.}, archive = {J_EAAI}, author = {Shun Xu and Sheng Cheng and Shuyang Jin and Xudong Hu and Weitao Wu and Zhong Xiang}, doi = {10.1016/j.engappai.2025.112296}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112296}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Industrial fabric defect-generative adversarial network (IFD-GAN): High-fidelity fabric cross-scale defect samples synthesis method for enhancing automated recognition performance}, volume = {161}, year = {2025}, }
• (2025). Sample-efficient planning-control framework for autonomous underwater vehicle docking using lightweight cross Q-learning. EAAI, 161, 112293. (www)

  @article{ , abstract = {Autonomous Underwater Vehicles (AUVs) are essential for long-duration missions in underwater sensor networks, but their operation is constrained by limited onboard energy and low-bandwidth communication. Docking stations are deployed to support energy replenishment and data transfer, requiring AUVs to autonomously plan return trajectories and achieve precise docking. Traditional motion planning and control methods for AUV docking often decouple global motion planning from local control or rely on accurate dynamics models, while existing deep reinforcement learning (DRL) approaches frequently neglect sample efficiency. This work presents a novel DRL framework that tightly integrates motion planning and control while emphasizing sample efficiency to enable reliable and efficient AUV docking. The docking task is formulated as an Markov Decision Process (MDP), in which a structured reward function is defined to guide the agent towards goal-directed behavior in an interpretable and sample-efficient manner. A lightweight Cross Q-Learning (CrossQ-Lite) algorithm is developed by removing target networks and incorporating batch normalization, achieving stable learning with approximately 92.8% fewer parameters while maintaining comparable performance. Simulation results demonstrate that, under the proposed MDP formulation, CrossQ-Lite achieves a minimum of 1 . 39 × higher sample efficiency and improved training stability compared to state-of-the-art DRL baselines. The trained model reliably completes AUV docking tasks, providing a practical and high-performance solution for autonomous underwater sensor networks.}, archive = {J_EAAI}, author = {Kaixin Zhang and Yuelong Xu and Minghao Zhao and Yu Jiang}, doi = {10.1016/j.engappai.2025.112293}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112293}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Sample-efficient planning-control framework for autonomous underwater vehicle docking using lightweight cross Q-learning}, volume = {161}, year = {2025}, }
• (2025). Fault diagnosis of wind turbine based on dual-channel feature aggregation network with attentional mechanism. EAAI, 161, 112291. (www)

  @article{ , abstract = {Wind turbines are subjected to alternating stresses and shock loads during operation, and the collected vibration signals are nonlinear, non-stationary and contain noise, resulting in subtle fault features that are difficult to extract. In order to extract discriminative features from vibration signals under variable speed operating conditions, a dual-channel feature aggregation network (DCNet) with attention mechanism is developed in this paper. First, a Parallel Patch-Aware Convolutional Module (PPCM) is constructed to extract feature information of different scales and levels from Time–Frequency Representations (TFR). Specifically, convolutional operations are performed in both the spatial and frequency domains, endowing it with local–global capturing capabilities and efficiency. Then, to improve the network operation rate, Haar Wavelet Downsampling (HWD) is embedded into the DCNet architecture. The core idea is to reduce the spatial resolution of features through Haar wavelet transform while preserving as many discriminative features as possible. Additionally, Channel Prior Convolutional Attention (CPCA) is introduced to enable DCNet to focus on more critical feature parts by dynamically allocating channel and spatial attention weights, thereby suppressing redundant feature interference. Finally, in the Deep Feature Fusion Module (DFFM), a cross-attention mechanism is adopted for global interaction, fusing features from different channels to achieve feature enhancement. To obtain good diagnostic results under variable speed conditions, we apply a label smoothing algorithm to assist model training. The experimental results show that the proposed model has high diagnostic accuracy, can generalize effectively, and it still maintains high diagnostic accuracy under noise and variable speed working conditions.}, archive = {J_EAAI}, author = {Haiyu Guo and Xingzheng Guo and Xiaoguang Zhang and Fanfan Lu and Chuang Liang}, doi = {10.1016/j.engappai.2025.112291}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112291}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fault diagnosis of wind turbine based on dual-channel feature aggregation network with attentional mechanism}, volume = {161}, year = {2025}, }
• (2025). A new deep learning framework for intelligent aerial monitoring of power transmission line insulators. EAAI, 161, 112290. (www)

  @article{ , abstract = {Transmission line insulators are critical components for maintaining the integrity and efficiency of power transmission lines. Monitoring the health, performance, and efficiency of this network has always been of interest since the past. Traditional and field inspection methods are often tedious, costly, and prone to human error. In order to monitor the condition of transmission lines and overcome traditional challenges, this study proposes an intelligent monitoring framework using unmanned aerial vehicle (UAV) imaging and advanced deep learning models to identify and classify insulator defects. The dataset used in this study consists of 870 original images, 600 healthy images, 140 damaged images, and 130 images of flashover insulators collected from UAV flights and online sources. To increase the diversity of the training data, a data augmentation technique was used, increasing the number of images to 1952 samples and dividing the data into 80 % training sets and 20 % test sets. While the current evaluation is performed on this dataset, real-world experiments will be considered in future studies to further validate the model. In this approach, three scenarios were considered to evaluate and detect these errors. The first scenario used an independent convolutional neural network (CNN) for feature extraction and classification, achieving an accuracy of 82.54 %. The second scenario combined CNN with transfer learning (TL) techniques to improve feature representation, achieving an accuracy of 95.99 %. Scenario Three integrated a hybrid CNN model with a long short-term memory network, achieving an average accuracy of 99.67 % in fault detection and classification. The results demonstrate the superiority and robustness of the proposed hybrid model, providing a reliable and cost-effective solution for transmission line insulator monitoring. This study highlights the significant potential of combining UAV imagery with deep learning algorithms, reassuring the audience of the effectiveness of the proposed solution.}, archive = {J_EAAI}, author = {Reza Rahimi Nejadbougar and Ebadat Ghanbari Parmehr and Alireza Afary and Samira Mavaddati}, doi = {10.1016/j.engappai.2025.112290}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112290}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A new deep learning framework for intelligent aerial monitoring of power transmission line insulators}, volume = {161}, year = {2025}, }
• (2025). Spatial-temporal generative network based on deep long short-term memory autoencoder for hand skeleton data sequences reconstruction and recognition. EAAI, 161, 112289. (www)

  @article{ , abstract = {Convolutional neural networks attract the highest research focus in the developing field of Hand Gesture Recognition (HGR). Nevertheless, these approaches presented a challenging task in adapting to time-series data. In skeleton-based HGR, extracting spatial–temporal information remains a challenge. In recent times, recurrent neural networks have exhibited exceptional performance in detecting desired hand gestures by processing of varied length time-series data. Although they outperform traditional methods when huge training data is accessible, their effectiveness significantly diminishes when data availability is constrained. In this study, we introduce an unsupervised data augmentation network known as the Spatial-Temporal Generative Network (STGN), which reconstructs both the spatial and temporal information of the input sequences by leveraging a Deep Long Short-Term Memory Auto-Encoder (DLSTM-AE) network. Consequently, the DLSTM-AE combined with different Long Short-Term Memory (LSTM) network variations, forming an integrated network that can be trained end-to-end for HGR. Through experimentation conducted on the LeapGestureDB dataset (Leap Motion-based Gesture Dataset) and RIT dataset (Rochester Institute of Technology Hand Gesture Dataset), we prove that data reconstruction using STGN had a prominent effect on improving the accuracy of recognizing time-series based hand gestures. For all experiments, the best recognition results are achieved in the augmented dataset. Accuracies were improved on all tested LSTM networks from 2 to 10%. For reproducible research, the code is available at: https://github.com/AMEURsafa/STGN .}, archive = {J_EAAI}, author = {Safa Ameur and Mohamed Ali Mahjoub and Anouar Ben Khalifa}, doi = {10.1016/j.engappai.2025.112289}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112289}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Spatial-temporal generative network based on deep long short-term memory autoencoder for hand skeleton data sequences reconstruction and recognition}, volume = {161}, year = {2025}, }
• (2025). Neural dynamic fluid reconstruction technique for four-dimensional imaging of combustion flame based on deep learning. EAAI, 161, 112288. (www)

  @article{ , abstract = {Three-dimensional optical diagnostic techniques based on the principles of tomographic imaging enable the acquisition of rich three-dimensional information in experimental flow fields through reconstruction calculations. However, for tasks involving the reconstruction of three-dimensional flow fields at high temporal resolutions, existing methods incur high computational costs, low reconstruction efficiency, and struggle to achieve high spatiotemporal resolution measurements. This paper proposes the Neural Dynamic Fluid Reconstruction Technique (NDFRT) based on deep learning. NDFRT incorporates the time dimension into the reconstruction scope to achieve the four-dimensional reconstruction of dynamic flow fields using neural networks. NDFRT has the following technical advantages: (1) ultra-high spatiotemporal reconstruction resolution; (2) good computational efficiency, with the reconstruction parameter scale only half that of traditional voxel-based methods; (3) the ability to perform three-dimensional frame prediction of dynamic fluids. We validated the proposed method using numerical simulation and experimental jet flame reconstruction and compared it with the traditional algebraic reconstruction technique (ART). Experimental results demonstrate that NDFRT outperforms traditional ART methods in terms of computational efficiency, reconstruction resolution, and reconstruction accuracy.}, archive = {J_EAAI}, author = {Fuhao Zhang and Zhiyin Ma and Can Gao and Gang Xun and Qingchun Lei and Xuesong Li}, doi = {10.1016/j.engappai.2025.112288}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112288}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Neural dynamic fluid reconstruction technique for four-dimensional imaging of combustion flame based on deep learning}, volume = {161}, year = {2025}, }
• (2025). BeltDiff: Diffusion-based self-labeled generation system for conveyor belt damage detection. EAAI, 161, 112287. (www)

  @article{ , abstract = {The damages of conveyor belt are abnormal occurrence of industrial production line, generally resulting in serious economic losses. While utilizing deep-learning-based methods in belt damage detection task, models training is challenging due to the prevalent scarcity of both numbers and diversity in dataset, leading to the restricted performance and limited accuracy in practical applications. To address this issue, we propose a novel self-labeled generation system based on diffusion model, for creating effective and large-scale conveyor belt detection datasets. Specifically, the training of our system is divided into three stages for progressive task decomposition and interpretable user control. Among the generation process, we focus on the textual prompt alignment and shape guidance to generate more reasonable and high-quality damaged belt images. Meanwhile, larger scale datasets are created via our method on two scenes, including base scene with limited samples, and brand new scenes with no damaged samples, which compose 9636 and 7500 images respectively. By evaluating the quality of damaged images, and the relations between generated and real images, our method demonstrates its effectiveness on generating damaged belt image pairs with annotations. Further, for verifying the validity of our generative datasets, we implement numerous experiments of 4 types of popular detection models training on different settings. The results demonstrate that our datasets support effective accuracy improvement, comparing with the base datasets and the classic augmented datasets, specifically increased by up to 14.4% in the brand new scenes with no damaged samples.}, archive = {J_EAAI}, author = {Peixian Zhuang and Yuanxiu Cai and Xi Liu and Xianchao Zheng and Fuheng Xiao and Jiangyun Li}, doi = {10.1016/j.engappai.2025.112287}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112287}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {BeltDiff: Diffusion-based self-labeled generation system for conveyor belt damage detection}, volume = {161}, year = {2025}, }
• (2025). A robust deep learning based model for denoising phonocardiogram signals in clinical environments. EAAI, 161, 112286. (www)

  @article{ , abstract = {A cardiac auscultation, being a very popular first line cardiovascular diseases screening method, typically produces phonocardiogram signals contaminated by background sounds, e.g., speech and movement of other patients and doctors/nurses, phone ringing and door knocking/opening/closing, and internal physiological body noises, e.g., lung sounds, muscle contraction and motion artifacts, as it is usually done in a noisy environment and it is not possible to suppress all the internal physiological body noises. So, a quality of the recording is finally decreased. Moreover, this can strongly affect a subsequent analysis/classification of the captured recording and can also lead to its incorrect classification. In this paper, we propose a novel deep convolutional architecture that integrates the bidirectional long short-term memory layers with the transformer blocks involving multi-head attention mechanism, which is particularly designed for denoising phonocardiogram signals. The performance of the proposed artificial intelligence-based approach is evaluated through synthetic data generated by using one public phonocardiogram dataset and two public real-world hospital ambient/respiratory sound databases as well as by deploying two public and renowned real-world clinical environment datasets containing phonocardiogram signals contaminated by various environmental and physiological noises recorded during the clinical auscultations. The deployed complex evaluation strategy assures a higher robustness and generalizability of the proposed approach. As also a result of that, the proposed model outperforms the current state-of-the-art approaches and provides the significant improvements in all the assessed signal quality and noise suppression metrics and over all the involved benchmark datasets. Finally, it should be noted here that the model achieved an average estimated SNR of 14.984 dB and 11.924 dB for the PASCAL and CirCor22 dataset respectively, which represent the renowned real-world clinical environment datasets.}, archive = {J_EAAI}, author = {Maros Jakubec and Eva Lieskovska and Peter Pocta}, doi = {10.1016/j.engappai.2025.112286}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112286}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A robust deep learning based model for denoising phonocardiogram signals in clinical environments}, volume = {161}, year = {2025}, }
• (2025). Transformer-based end-to-end point restoration network for filtering temporarily-static objects from point cloud. EAAI, 161, 112284. (www)

  @article{ , abstract = {The precision of map-based localization systems is pivotal for advanced navigation and autonomous vehicle technologies. This study addresses the challenge of dynamic environments where temporarily-static objects, such as vehicles, introduce occlusion and temporal alterations to the mapping landscape, leading to inaccuracies in mapping and localization. We propose a transformer-based deep learning model that removes temporarily-static objects and refines occluded areas in raw point cloud data, thereby enhancing both mapping and localization. Our method leverages a transformer-based point restoration network trained on a novel dataset that does not require human-driven manual annotation and predicts point clouds free of temporarily-static objects. The approach was validated through experiments demonstrating improved qualitative and quantitative results. The results display the importance of our method by removing temporarily-static objects and refining occluded areas. The proposed method achieves a 26.4 % reduction in Chamfer Distance L1 (CD-L1), 33.1 % reduction in Chamfer Distance L2 (CD-L2), and a 30.2 % relative improvement in the harmonic mean of precision and recall (F-Score) over the baseline network, demonstrating significantly enhanced model for filtering. This research contributes to artificial intelligence (AI)-enabled autonomous navigation by offering a systematic solution to a common problem in dynamic environments, setting a new filtering technology for mapping systems.}, archive = {J_EAAI}, author = {Sabir Hossain and Xianke Lin}, doi = {10.1016/j.engappai.2025.112284}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112284}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Transformer-based end-to-end point restoration network for filtering temporarily-static objects from point cloud}, volume = {161}, year = {2025}, }
• (2025). Cross-view gait recognition based on discriminative global–local feature representation and learning. EAAI, 161, 112282. (www)

  @article{ , abstract = {Performance of gait recognition can be heavily influenced by deformation of walking silhouettes due to variations of view angle. Cross-view gait recognition is still one of the most challenging problems in pattern recognition community. In this paper, we propose a new cross-view gait recognition algorithm by discriminative global–local feature representation and learning. First, we design cross-view convolutional neural network for learning global gait feature representations underlying binary walking silhouettes sequences. Second, we propose to extract local gait features by partial spatial–temporal sequence and local feature extractors, which can represent the motion characteristics and temporal changes of different human body parts. Both global and local features in our proposed method are extracted from different gait representations to provide spatial–temporal information from diverse aspects, and they are fused on feature level to further reduce the sensitivity to variations of view angle and improve the performance of gait recognition. Finally, we introduce deep metric learning framework and propose a robust, effective, and view-related loss function, named view triplet loss function, to learn discriminative gait features. Compared with triplet loss, the proposed loss function explores hierarchical relationship of the variations of view angles, which achieves better intra-subject compactness and inter-subject separability. Experimental results on CASIA-B, OULP and OUMVLP gait datasets demonstrate the feasibility and advantage of our proposed algorithm.}, archive = {J_EAAI}, author = {Muqing Deng and Yi Zou and Zhi Zeng and Xiaoreng Feng and Yanjiao Wang and Yuan Liu}, doi = {10.1016/j.engappai.2025.112282}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112282}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Cross-view gait recognition based on discriminative global–local feature representation and learning}, volume = {161}, year = {2025}, }
• (2025). A novel clustering-ensemble learning model for day-ahead photovoltaic power forecasting. EAAI, 161, 112281. (www)

  @article{ , abstract = {Accurate day-ahead photovoltaic power forecasting (PPF) is essential for grid scheduling and transaction planning. However, time series prediction models based on historical meteorological and photovoltaic power data often struggle to capture long-term dependencies. Existing research typically employs multivariate regression models to establish mapping relationships between numerical weather prediction (NWP) and photovoltaic power data, yet still exhibits deficiencies in weather pattern recognition and consideration of temporal patterns. Therefore, this paper proposes a novel clustering-ensemble learning model to enhance predictive performance. In the proposed deep time-series clustering algorithm, key features from NWP data are extracted for clustering, addressing the inefficiencies of traditional clustering methods in handling high-dimensional data. A weighted Warp-Euclidean distance is proposed to capture sequence trend homogeneity and spatial similarity, overcoming the limitations of Euclidean distance in time-series data. The ensemble learning model combines the advantages of two gradient boosting tree models to handle nonlinear features and high-dimensional data, while introducing lag features to enhance temporal dynamic learning capability. This approach enhances the stability and mitigates the limitations associated with relying primarily on weather-related features. This study utilizes NWP and measured photovoltaic data from two stations in Hebei Province, China. Results demonstrate superior performance compared to other clustering algorithms and prediction models in day-ahead PPF tasks, achieving R 2 scores of 0.952 and 0.943 on two public photovoltaic datasets, representing improvements of 0.020 and 0.017 over optimal benchmark models. The proposed framework provides theoretical guidance for the stable operation of grid-connected photovoltaic systems.}, archive = {J_EAAI}, author = {Feifei Yang and Xueqian Fu and Zhengshuo Li and Dawei Qiu and Hamed Badihi}, doi = {10.1016/j.engappai.2025.112281}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112281}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel clustering-ensemble learning model for day-ahead photovoltaic power forecasting}, volume = {161}, year = {2025}, }
• (2025). Artificial intelligence-based phonocardiogram signal classification using segment-specific multi-domain features for cardiovascular and arterial disease. EAAI, 161, 112280. (www)

  @article{ , abstract = {Heart diseases are the leading cause of death globally, and early diagnosis coupled with timely treatment can save lives. The human heart produces sounds that reflect the functions of its valves and chambers, with murmurs often indicating abnormal heart behavior. The phonocardiogram (PCG) analysis offers a quick, easy, and affordable diagnostic method for detecting such conditions. The PhysioNet/2016 dataset, containing annotated heart sound recordings, has advanced automated heart sound classification research. Despite the dataset encompassing nine classes of heart diseases, it has primarily been used for binary classification tasks. Multi-class classification is challenging due to overlapping acoustic features among certain heart conditions and signal variability. To address these issues, we implemented segment specific feature extraction across six domains — time, frequency, amplitude, wavelet, spectrum, signal processing — and evaluated various machine learning classifiers. Our findings revealed that the Random Forest classifier outperformed others, leading us to integrate it into our model. On the testing dataset, it attained 94% accuracy, 93% sensitivity and 93% specificity for binary classification Our model achieved 89% accuracy, 80% sensitivity and 98% specificity for nine-class classification including both valve diseases and artery-related conditions.}, archive = {J_EAAI}, author = {Syeda Sana Bukhari and Shahab U. Ansari and Khurram Khan Jadoon and Raja Hashim Ali}, doi = {10.1016/j.engappai.2025.112280}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112280}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Artificial intelligence-based phonocardiogram signal classification using segment-specific multi-domain features for cardiovascular and arterial disease}, volume = {161}, year = {2025}, }
• (2025). Probabilistic forecasting of non-ferrous metal prices based on outlier treatment algorithms, quantile regression based deep learning and two-phase multi-objective optimization. EAAI, 161, 112277. (www)

  @article{ , abstract = {Accurate forecasting of non-ferrous metal prices plays a crucial role in helping enterprises optimize production, control costs, and manage risks, as well as in enabling countries and industries to formulate resource strategies, ensure economic security, and promote sustainable development. However, existing research still has limitations in outlier detection, uncertainty analysis, and model optimization, making it difficult to meet the forecasting needs in complex market environments. To fill these gaps, this study proposes a novel combined probabilistic forecasting framework for non-ferrous metal prices. First, this study utilizes outlier treatment algorithms to achieve robust outlier detection and correction. Second, multiple quantile regression-based deep learning models are established for probabilistic forecasting. Subsequently, a two-phase multi-objective optimization strategy is proposed. Specifically, the first phase is designed to optimize the upper quantile half-interval and lower quantile half-interval of each quantile regression-based deep learning model, while the second phase is used to optimize the combined weights of the aforementioned single model to synergistically enhance interval reliability and prediction accuracy. Finally, silver and aluminum futures prices are used as illustrative case studies. The experimental results show that the proposed model outperforms the benchmark method in terms of interval coverage and prediction accuracy. This study provides a scientific basis and reference for industrial production optimization, investment decision-making and resource management planning.}, archive = {J_EAAI}, author = {Pei Du and Fangrui Gui and Lijie Shan and Qianyi Xing and Jianzhou Wang}, doi = {10.1016/j.engappai.2025.112277}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112277}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Probabilistic forecasting of non-ferrous metal prices based on outlier treatment algorithms, quantile regression based deep learning and two-phase multi-objective optimization}, volume = {161}, year = {2025}, }
• (2025). Identifying nonlinear roll damping and restoring parameters via physics-informed neural network. EAAI, 161, 112275. (www)

  @article{ , abstract = {To effectively characterize a ship’s roll motion, developing a robust dynamic model that incorporates accurate damping and restoring parameters is essential. Given that only limited free-decay measurements have been measured, an efficient physics-informed neural network-based approach has been developed to simultaneously determine the nonlinear damping and restoring parameters associated with ship rolling. Rather than relying on a large dataset to train a neural network for parameter identification, the proposed method incorporates the physical equations of roll motion into the residual networks, ensuring that the identified parameters adhere to their physical interpretation. Three numerical cases, encompassing computational fluid dynamics (CFD) analyses, and laboratory experiments, are presented to show the superior performance of the developed method. Key findings indicate that the proposed method effectively identifies the investigated parameters with high precision while maintaining consistent performance across various initial roll angle conditions. Compared to traditional methods, the proposed approach offers significant advantages, including reduced reliance on large datasets, automated parameter identification without manual tuning, and the ability to incorporate physical constraints directly into the learning process. These features make the method particularly suitable for real-world applications where data is sparse or costly to obtain.}, archive = {J_EAAI}, author = {Shuai Cong and Jinwei Sun and Qianying Cao and Changhong Zhi and Shixuan Liu}, doi = {10.1016/j.engappai.2025.112275}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112275}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Identifying nonlinear roll damping and restoring parameters via physics-informed neural network}, volume = {161}, year = {2025}, }
• (2025). Anomaly detection in offshore open radio access network using long short-term memory models on a novel artificial intelligence-driven cloud-native data platform. EAAI, 161, 112274. (www)

  @article{ , abstract = {The Radio Access Network (RAN) is a critical component of modern telecommunications infrastructure, currently evolving towards disaggregated and open architectures. These advancements are pivotal for integrating intelligent, data-driven applications aimed at enhancing network reliability and operational autonomy through the introduction of cognitive capabilities, as exemplified by the emerging Open Radio Access Network (O-RAN) standards. Despite its potential, the nascent nature of O-RAN technology presents challenges, primarily due to the absence of mature operational standards. This complicates the management of data and intelligent applications, particularly when integrating with traditional network management and operational support systems. Divergent vendor-specific design approaches further hinder migration and limit solution reusability. These challenges are compounded by a skills gap in telecommunications business-oriented engineering, which remains a key barrier to effective O-RAN deployment and intelligent application development. To address these challenges, Boldyn Networks developed a novel cloud-native data analytics platform, specifically designed to support scalable Artificial Intelligence (AI) integration within O-RAN deployments. This platform underwent rigorous testing in real-world scenarios, and applied advanced AI techniques to improve operational efficiency and customer experience. Implementation involved adopting Development Operations (DevOps) practices, leveraging data lakehouse architectures tailored for AI applications, and employing sophisticated data engineering strategies. The platform successfully addresses connectivity challenges inherent in real-world offshore wind farm deployments using Long Short-Term Memory (LSTM) models for anomaly detection in network connectivity. After integrating the LSTM models into the network control, more than 90 percent of connectivity issues were reduced in runtime. This marks a step toward autonomous, self-organizing, and self-healing networks.}, archive = {J_EAAI}, author = {Abdelrahim Ahmad and Peizheng Li and Robert Piechocki and Rui Inacio}, doi = {10.1016/j.engappai.2025.112274}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112274}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Anomaly detection in offshore open radio access network using long short-term memory models on a novel artificial intelligence-driven cloud-native data platform}, volume = {161}, year = {2025}, }
• (2025). Structural design through reinforcement learning. EAAI, 161, 112273. (www)

  @article{ , abstract = {This paper introduces the Structural Optimization gym (SOgym), a novel open-source Reinforcement Learning (RL) environment designed to advance machine learning in Topology Optimization (TO). SOgym enables RL agents to generate physically viable and structurally robust designs by integrating the physics of TO into the reward function. To enhance scalability, SOgym leverages feature-mapping methods as a mesh-independent interface between the environment and the agent, allowing efficient interaction with the design variables regardless of mesh resolution. Baseline results use a model-free Proximal Policy Optimization agent and a model-based DreamerV3 agent. Three observation space configurations were tested. The Topology Optimization game ( TopOpt game) inspired configuration, an interactive educational tool for designing structures to minimize compliance under volume constraints, performed best in terms of performance and sample efficiency. The 100 million parameter DreamerV3 model (DreamerV3-100M) produced structures with compliance values approximately 54 % higher than those from traditional optimization methods and a 0 % disconnection rate, an improvement over supervised learning approaches that often struggle with disconnected load paths. When comparing the learning rates of the agents to those of engineering students from the TopOpt game experiment, the DreamerV3-100M model shows a learning rate approximately four orders of magnitude lower, an impressive feat for a policy trained from scratch through trial and error. These results suggest RL's potential to solve continuous TO problems and its capacity to explore and learn from diverse design solutions. SOgym provides a platform for developing RL agents for structural design challenges and is publicly available to support research in the field.}, archive = {J_EAAI}, author = {Thomas Rochefort-Beaudoin and Aurelian Vadean and Niels Aage and Sofiane Achiche}, doi = {10.1016/j.engappai.2025.112273}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112273}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Structural design through reinforcement learning}, volume = {161}, year = {2025}, }
• (2025). Single-channel selection for epileptic seizure identification through a custom machine learning model. EAAI, 161, 112272. (www)

  @article{ , abstract = {Epilepsy affects millions of people worldwide, triggering undesirable motor and sensory events that drastically impair the quality of life of those affected. Detecting this condition through a monitoring system can be extremely useful in identifying epileptic seizures without relying exclusively on a specialist. Our goal is to identify these seizures using personalized supervised machine learning models – eXtreme Gradient Boosting – for each patient, employing only one electroencephalogram input channel. Our methodology explores the relationship between the selected input channel and the topographic location of epileptic seizures modeled for the patient. The results revealed notable accuracy performances for the three patients investigated: 100%, 99%, and 88%, after applying the proposed time filter that checks for the presence/absence of a seizure every 3 s. Furthermore, it was possible to observe that the results are consistent with the affected area in each patient, demonstrating the effectiveness of the method in selecting the single channel. This approach demonstrates the feasibility of detecting convulsive seizures using only one input channel and underscores the need to extend the methodology to more patients.}, archive = {J_EAAI}, author = {Jusciaane Chacon Vieira and Ignacio Sanchez-Gendriz and Luiz Affonso Guedes}, doi = {10.1016/j.engappai.2025.112272}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112272}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Single-channel selection for epileptic seizure identification through a custom machine learning model}, volume = {161}, year = {2025}, }
• (2025). IMobileTransformer: A fusion-based lightweight model for rice disease identification. EAAI, 161, 112271. (www)

  @article{ , abstract = {Rice blast, sheath blight, leaf scald, bacterial leaf blight, and brown spot severely threaten rice yield. To address the limitations of current deep learning methods in rice disease recognition, particularly their insufficient integration of local and global features, this study proposes an Improved MobileTransformer (IMobileTransformer) model. The proposed architecture synergistically combines MobileNet’s strengths in local feature extraction and lightweight architecture with Transformer’s superior capability in global information processing. Specifically, the model is designed with three functional branches: a) a MobileNet branch utilizing inverted residual structure with depthwise separable convolution layers to reduce parameters and computational complexity, b) a Transformer branch modified from Swin-Transformer architecture, where the Multilayer Perceptron (MLP) layer is enhanced by splitting input feature channels through an Inception-based structure to maintain global feature extraction efficiency while minimizing computational overhead, and c) a feature fusion branch that concatenates reshaped outputs from both branches through channel-wise stacking, enabling effective integration of local and global representations. Experimental results show that compared to classical models such as MobileNetV3-Large, EfficientNet-B0, Vision Transformer Base/16 (ViT-B/16), Shifted Window Transformer (Swin-Transformer), Tiny Vision Transformer (TinyViT), Mobile Vision Transformer (MobileViT), LocalViT-S, IMobileTransformer achieves a recognition accuracy of 99.62% for rice diseases, with improvements of 1.71%, 0.91%, 38.09%, 4.17%, 1.99%, 1.5% and 0.42%, respectively, providing an effective solution for rice disease recognition.}, archive = {J_EAAI}, author = {Yang Lu and Haoyang Zhou and Peng Wang and Erzhi Wang and Gongfa Li and Tongjian Yu}, doi = {10.1016/j.engappai.2025.112271}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112271}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {IMobileTransformer: A fusion-based lightweight model for rice disease identification}, volume = {161}, year = {2025}, }
• (2025). Personalized and safe medication recommendation based on convolutional neural network and transformer architecture. EAAI, 161, 112267. (www)

  @article{ , abstract = {With the accumulation of electronic health records (EHRs), artificial intelligence (AI) based medical services such as medication recommendation (MR) has aroused widespread concern. However, existing drug recommendation models suffer from inadequate patient representation and adverse drug–drug interactions (DDIs). To address these challenges, we propose an AI-based personalized and safe medication recommendation method based on convolutional neural network and transformer architecture (CT-PASMR). Specifically, convolutional neural network (CNN) and transformer are combined in parallel (CAT) to model local relationships in a patient’s single visit and long-term dependencies in sequential EHR data, respectively. Subsequently, graph attention networks (GATs) are deployed to capture drug co-occurrences and adverse DDIs with various weights, generating safe drug representations. Moreover, a joint loss function is introduced to balance accuracy and safety in CT-PASMR. Finally, the experimental results on MIMIC (Medical Information Mart for Intensive Care)-III and MIMIC-IV datasets demonstrate that CT-PASMR achieves competitive performance on seven evaluation metrics such as DDI rate, Jaccard index and F1 score, compared with nine state-of-the-art (SOTA) baseline models. Ultimately, ablation studies and further analysis confirm the efficacy of CATs and GATs in providing personalized and safe medication recommendations. Code, qualitative results, and trained weights will be available at the link: https://github.com/gefengru/CT-PASMR .}, archive = {J_EAAI}, author = {Fengru Ge and Xiaomei Yu and Xue Li and Xingxu Fan and Yanjie Zhao}, doi = {10.1016/j.engappai.2025.112267}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112267}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Personalized and safe medication recommendation based on convolutional neural network and transformer architecture}, volume = {161}, year = {2025}, }
• (2025). Contrastive learning with similarity enhancement for dimensionality reduction. EAAI, 161, 112266. (www)

  @article{ , abstract = {Dimensionality reduction aims to reduce the number of dimensions while preserving crucial information. As a self-supervised learning approach, contrastive learning provides a novel perspective for dimensionality reduction. However, most contrastive learning methods focus on optimizing similarity, which have limitations in reducing feature redundancy. Moreover, the dependence on negative pairs introduces computational overhead. To address these problems, we propose Contrastive Learning with Similarity Enhancement for Dimensionality Reduction (CLSDR), which integrates neighborhood embedding into the contrastive learning framework. Specifically, CLSDR uses k -nearest neighbors sampling to construct positive pairs. We design a multi-level loss function that captures the diversity of data while maintaining the consistency of local and global features. In addition, we propose a nonlinear similarity optimization mechanism based on logarithmic smoothing, which adjusts the gradient of similarity loss, improving the stability during the training process. Experimental results demonstrate that CLSDR significantly outperforms several state-of-the-art methods. Especially, on the Street View House Numbers dataset, CLSDR achieves 66.7% and 62.3% Top-1 accuracy with two classifiers in 64-dimensional embedding space, which have 10.6% and 38.4% improvement over the best competing approach. Thus, CLSDR exhibits strong robustness and scalability across different dimensions.}, archive = {J_EAAI}, author = {Qi Yang and Changpeng Wang and Linlin Feng and Lizhen Ji and Jiangshe Zhang}, doi = {10.1016/j.engappai.2025.112266}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112266}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Contrastive learning with similarity enhancement for dimensionality reduction}, volume = {161}, year = {2025}, }
• (2025). A multi-modal approach for detecting drivers’ distraction using bio-signal and vision sensor fusion in driver monitoring systems. EAAI, 161, 112265. (www)

  @article{ , abstract = {According to a report by the World Health Organization (WHO), approximately 1.3 million people lose their lives annually owing to traffic accidents. The majority of road traffic accidents stem from driver negligence. Recently, there has been a growing interest in utilizing deep learning and machine learning technologies to enhance the safety and efficiency of road traffic, with the aim of addressing issues arising from driver inattentiveness. Most studies focus on detecting abnormal driver behavior using driving sensors or driver images; however, they often overlook physiological factors such as the driver's bio-signals. Considering that the driver's state, including fatigue, stress, and concentration, can significantly affect driving safety, it is crucial to build models that consider biometric information. Therefore, this study proposes a multi-modal transformer model called Bio-Vision Transformer (BiViT) that comprehensively considers both driver bio-signals and images. The BiViT model uses a vision transformer to extract features from driver images and employs a time-series transformer to capture features from the driver's bio-signals. In addition, the interactions between the extracted features are modeled, and the joint fusion method is employed as the feature-fusion approach. To validate the proposed model, performance comparisons and analyses were conducted using commonly used models in image analysis. The experimental results demonstrated that the proposed BiViT model exhibited high performance, with an accuracy of 0.91 and a harmonic mean of precision and recall (F1-score) of 0.91, surpassing the performance of the comparison models.}, archive = {J_EAAI}, author = {Byeongjoon Noh and Myeongseok Park and Yechan Han and Jaeyun Kim}, doi = {10.1016/j.engappai.2025.112265}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112265}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-modal approach for detecting drivers’ distraction using bio-signal and vision sensor fusion in driver monitoring systems}, volume = {161}, year = {2025}, }
• (2025). Graph neural network with generative adversarial training for node classification on class imbalanced data. EAAI, 161, 112264. (www)

  @article{ , abstract = {Node classification in class-imbalanced graph data remains a critical challenge, as traditional graph neural networks (GNNs) either assume class-balanced graph structures or inadequately address class imbalance. This often results in predictive bias, where majority classes are favored while minority classes are underrepresented. To overcome this limitation, this study introduces a novel graph neural network with generative adversarial training (GNN-GAN), where the GNN extracts latent features from input node attributes balanced through data synthesis using a conditional generative adversarial network (GAN) and data fusion strategy. The GNN and GAN are trained synchronously to ensure GAN synthesizes samples that match real data distribution, while GNN adjusts to the quality of synthesized data in a timely manner. A data fusion strategy combines synthetic and real samples to mitigate class imbalance and maintain classification accuracy. Experiments on several benchmark graph datasets demonstrate that the GNN-GAN consistently outperforms state-of-the-art baselines. A comprehensive ablative study further validates the advantages of the synchronized training procedure, offering insights into the model's robustness across graph datasets with varying structures and imbalance ratios.}, archive = {J_EAAI}, author = {Xiaoqi Zhou and Peixin Shi}, doi = {10.1016/j.engappai.2025.112264}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112264}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Graph neural network with generative adversarial training for node classification on class imbalanced data}, volume = {161}, year = {2025}, }
• (2025). Diffusion-based feature absorption approach for improving lightweight object detectors in adverse weather conditions. EAAI, 161, 112263. (www)

  @article{ , abstract = {Although cutting-edge lightweight detectors have achieved desirable performance in favorable weather conditions, they again failed to accurately identify objects within low-quality scenes captured in inclement environments, especially during rainy nighttime. These detectors face difficulties learning beneficial information for detection because objects are obscured by rain and low-light conditions. To address the aforementioned challenges, we introduce an innovative and effective Diffusion-based Feature Absorption Learning (Diff-FAL) approach, to reinforce the performance of lightweight detection models in nighttime with rain interference. By developing an unsupervised training strategy based on the diffusion model, the proposed approach assists lightweight detectors absorb useful features from degraded images. To this end, our Diff-FAL framework comprises three distinct subnetworks: a Feature Optimization (FO) subnetwork, a Feature Mutation (FM) subnetwork, and a Lightweight Detection (LD) subnetwork. The FO subnetwork is designed to produce sharp, detailed features from rainy night images and the FM subnetwork is developed to transfer those features to the LD subnetwork. Comprehensive experiments on various datasets confirmed the superiority of our model, outperforming the compared method by up to 28.01% and 30.16% in accuracy on two published datasets: the rainy nighttime (RNT) and real rainy images (rRain) datasets, respectively, while maintaining high-speed performance during inference.}, archive = {J_EAAI}, author = {Quoc-Viet Hoang and Trung-Hieu Le}, doi = {10.1016/j.engappai.2025.112263}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112263}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Diffusion-based feature absorption approach for improving lightweight object detectors in adverse weather conditions}, volume = {161}, year = {2025}, }
• (2025). A multi-modal model for removal of crack image shadow based on language prior information. EAAI, 161, 112262. (www)

  @article{ , abstract = {Computer-vision based crack detection highly depends on the quality of images and it remains to be a challenging task due to the effects of shadow. The utilization of deep learning algorithm, guided by prior information, has been demonstrated effective in the removal of shadows. However, it has limitations in handling images with shadowed cracks. A deep learning method incorporating language prior information and a multi-modal model is proposed in this study to improve the removal of shadow effects in restoring the crack images. Natural language processing (NLP) is used to extract the language prior information from image text description to enhance the crack recognition, and OTSU method is used to obtain the image prior information. A novel multi-modal model, named as the Text-based Shadow Removal Network (TSRNet), for the removal of shadow in crack image, is proposed to have better crack restoration capability. Bayesian optimization approach is also employed to optimize the network architecture improving the prediction precision of TSRNet. The proposed framework for shadow removal is verified using an open-source shadowed concrete crack image dataset and a new dataset from experiment. Results indicate that the language prior information can enhance the TSRNet model with better performances of shadow removal for real-world crack images compared with other existing models.}, archive = {J_EAAI}, author = {Gang Liu and Xuming Li and Qingshan Yang and S.S. Law and Changjun Deng}, doi = {10.1016/j.engappai.2025.112262}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112262}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-modal model for removal of crack image shadow based on language prior information}, volume = {161}, year = {2025}, }
• (2025). Development and implementation of a hybrid visual prediction algorithm for robotic smart tomato harvesting. EAAI, 161, 112261. (www)

  @article{ , abstract = {This paper introduces new algorithms designed specifically for smart tomato harvesting, which combines predictions from two independently trained YOLOv8 (You Only Look Once version 8) models, each specialized on different datasets to detect tomatoes and classify their ripeness stages—ripe, unripe, and semi-ripe—while also computing their center points. To enhance detection accuracy under varying field conditions, multiple fusion strategies were developed, including a hybrid algorithm that integrates union and confidence-weighted summation methods. The Hybrid Algorithm achieved the best performance, surpassing the original models and other fusion techniques. It attained F1 scores of 0.697 and 0.694 at confidence thresholds of 0.5 and 0.9, respectively, compared to the original models' F1 scores of 0.670 and 0.648 at confidence 0.5, and 0.481 and 0.497 at confidence 0.9. This corresponds to an improvement of 4.0 % and 7.6 % at confidence 0.5, and 44.3 % and 39.6 % at confidence 0.9, demonstrating the hybrid algorithm's stability and superiority, particularly at higher thresholds. Furthermore, the system utilizes a high-resolution RGB (Red, Green, Blue) camera for real-time image capture, enhancing model performance in complex agricultural environments. This study validates the effectiveness of confidence-based fusion in developing robust and accurate vision systems for precision agriculture.}, archive = {J_EAAI}, author = {Giuseppe Carbone and Angad Singh Gurtatta and Dmitry Malyshev}, doi = {10.1016/j.engappai.2025.112261}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112261}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Development and implementation of a hybrid visual prediction algorithm for robotic smart tomato harvesting}, volume = {161}, year = {2025}, }
• (2025). Multi-modal multi-level feature representation learning for flow pattern identification of oil-water two-phase flow. EAAI, 161, 112260. (www)

  @article{ , abstract = {To tackle the difficulty of the traditional experimental methods in real-time monitoring and identification of the flow process, this paper introduces a novel flow pattern identification method of the vertical oil-water two-phase flow based on multi-modal multi-level feature representation. The one-dimensional electromagnetic signals are encoded into two-dimensional feature spaces to explore their structural complexity, evolutionary probability laws and nonlinear characteristics in multimodal domain, thereby generating a multi-modal representation of the electromagnetic signals. Subsequently, a multi-modal multi-level feature fusion network is developed for flow pattern identification network, which flexibly leverages effective information across different modalities and levels, thereby enhancing the identifying accuracy. Experimental results demonstrate that the proposed method achieves high accuracy on the constructed multi-modal dataset, proving its feasibility and effectiveness in identifying the flow pattern of the oil-water two-phase flow in vertical pipes.}, archive = {J_EAAI}, author = {Weihang Kong and Yaohan Chi and He Liu and Hongbao Tang and He Li}, doi = {10.1016/j.engappai.2025.112260}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112260}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-modal multi-level feature representation learning for flow pattern identification of oil-water two-phase flow}, volume = {161}, year = {2025}, }
• (2025). Defining and evaluating decision and composite risk in language models applied to natural language inference. EAAI, 161, 112253. (www)

  @article{ , abstract = {Despite their impressive performance, large language models (LLMs) are known to pose important risks. One such set of risks arises from misplaced confidence, whether over-confidence or under-confidence, that the models have in their inference. While the former is well studied, the latter is not, leading to an asymmetry in understanding the comprehensive risk of the model based on misplaced confidence. In this paper, we address this asymmetry by defining two types of risk (decision and composite risk), and proposing an experimental framework consisting of a two-level inference architecture and appropriate metrics for measuring such risks in both discriminative and generative LLMs. The first level relies on a decision rule that determines whether the underlying language model should abstain from inference. The second level (which applies if the model does not abstain) is the model’s inference. This framework has direct implications for error-sensitive LLM-based engineering applications where reliable decision-making is critical, such as healthcare and finance. Through detailed experiments on four natural language commonsense reasoning datasets using both an open-source ensemble-based transformer model and a generative LLM, we demonstrate the practical utility of our evaluation framework. Our results show that the framework can get an LLM to confidently respond to an extra 20.1% of low-risk inference tasks that other methods might misclassify as high-risk, and skip 19.8% of high-risk tasks, which would have been answered incorrectly.}, archive = {J_EAAI}, author = {Ke Shen and Mayank Kejriwal}, doi = {10.1016/j.engappai.2025.112253}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112253}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Defining and evaluating decision and composite risk in language models applied to natural language inference}, volume = {161}, year = {2025}, }
• (2025). Advanced machine learning techniques for predicting wear performance in graphene oxide particulate interpenetrating polymer network composites. EAAI, 161, 112252. (www)

  @article{ , abstract = {This research investigates the wear behavior of hybrid polymeric composites made from synthetic glass and natural cotton fibers, reinforced with varying proportions of Graphene Oxide (GO) (0 %, 1 %, 3 %, 5 %, 7 %, 9 %). The effect of fiber arrangement and Graphene Oxide (GO) incorporation on wear rate and Coefficient of Friction (CoF) was evaluated using the Pin-On-Disk method, with analysis based on Taguchi's L 32 Orthogonal Array. The optimal parameters were found at 6 min, 5 % GO, 300 revolutions per minute (rpm) speed, 20 mm (mm) track diameter, and 10 N (N) load, achieving a minimum wear rate of 0.612 × 10 −4 cubic millimeters per newton-meter (mm 3 /N-m) and a CoF of 0.151. Predictive modeling was performed to predict the wear rate and coefficient of friction using supervised machine learning algorithms, including Linear Regression, Decision Tree, and Random Forest, to forecast material behavior. Performance evaluation using Confusion Matrix, Distribution Analysis, and various metrics showed that the Decision Tree model excelled, achieving near-perfect predictive power with a Mean Squared Error (MSE) of 0 and an R-squared value of 0.9999. The model demonstrated 100 % accuracy, with precision, recall, and F1-scores all equal to 1. This research demonstrates the effectiveness of combining natural and synthetic fibers with GO, along with the predictive power of machine learning in optimizing material properties.}, archive = {J_EAAI}, author = {Eastus Russel and S. Madhu and Judy S and Edwin Geo Varuvel and G.B. Santhi and G. Suresh and Femilda Josephin J.S and Mohammed F. Albeshr and Farzad Kiani}, doi = {10.1016/j.engappai.2025.112252}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112252}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Advanced machine learning techniques for predicting wear performance in graphene oxide particulate interpenetrating polymer network composites}, volume = {161}, year = {2025}, }
• (2025). Health state evaluation and analysis of equipment considering multi-scale data fusion. EAAI, 161, 112251. (www)

  @article{ , abstract = {Current researches of health state evaluation of equipment have increasingly emphasized the integration of data and knowledge to improve evaluation accuracy and interpretability, which leads to the emergence of hybrid information-based methods as a research hotspot. However, there are issues with redundant health indicators, multi-scale data with different features and measurement units, and interpretability measures in the evaluation. In this paper, an interpretable health state evaluation and analysis method of equipment is proposed based on trustworthy evidential reasoning rule (TERR). To deal with redundancy in health indicators, a multi-scale data analysis model is proposed based on the clustering analysis and Kruskal-Wallis test. It can select the optimal health indicators that effectively preserve the original evaluation information while reducing the model complexity. To integrate the selected indicators with uncertainty, a TERR-based multi-scale data fusion model is proposed, where the evidence weight, reliability, and trustworthiness are simultaneously considered. Also, an interpretable parameter optimization model is constructed to alleviate the uncertainty in initial evidential parameters. To study the interpretability of TERR, two sensitivity analysis methods are proposed, and the performance coefficient matrix of the model output to the perturbation is derived mathematically. This shows how the interpretability of TERR is enhanced and which external parameter has the greatest impact on outputs. Finally, a case study of health state evaluation of laser inertial measurement unit (LIMU) validates the effectiveness of the proposed method.}, archive = {J_EAAI}, author = {Shuai-Wen Tang and Jiang Jiang and Jian-Bin Sun and Zhuo-Ting Yu}, doi = {10.1016/j.engappai.2025.112251}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112251}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Health state evaluation and analysis of equipment considering multi-scale data fusion}, volume = {161}, year = {2025}, }
• (2025). Unmanned surface vehicle autonomous racing and obstacle avoidance with robust adversarial deep reinforcement learning. EAAI, 161, 112250. (www)

  @article{ , abstract = {This paper presents an autonomous racing control method for Unmanned Surface Vehicles (USVs) based on robust adversarial deep reinforcement learning (ADRL) algorithm, which leverages the strengths of both deep reinforcement learning and adversarial training. Adversarial obstacles and various tracks are employed to train a policy, so the proposed method can enhance the robustness and generalization of autonomous USV racing while ensuring effective obstacle avoidance. A simulation environment for USV racing was developed to conduct the experiments with Unity3D. The performance of the proposed method in handling diverse track scenarios and obstacle avoidance is demonstrated through simulations. Quantitative results show that ADRL achieves dramatic improvements over baseline DRL methods: collision rates are reduced by 99.85%, task completion rates are improved by 73.75%, and navigation time efficiency is enhanced by approximately 3% while maintaining superior safety performance.}, archive = {J_EAAI}, author = {Jingyu Liu and Yuanda Wang and Changyin Sun}, doi = {10.1016/j.engappai.2025.112250}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112250}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unmanned surface vehicle autonomous racing and obstacle avoidance with robust adversarial deep reinforcement learning}, volume = {161}, year = {2025}, }
• (2025). Multiscale constitutive modeling of anisotropic plasticity: Coupling the visco-plastic self-consistent model with the recurrent neural network and its implementation in finite element analysis. EAAI, 161, 112249. (www)

  @article{ , abstract = {The anisotropic and nonlinear strain-path-dependent nature of metal plasticity poses a major challenge for accurate constitutive modeling in finite element (FE) analysis. Traditional macroscale models are easily implemented but lack accuracy, while crystal plasticity (CP) models offer high fidelity at the cost of computational efficiency. To bridge this gap, we propose a deep neural network smart constitutive (DNNSC) framework that combines the visco-plastic self-consistent (VPSC) model with a gated recurrent unit (GRU) network. A VPSC model calibrated on pure aluminum generated 14,000 strain-paths for training GRU-based network. The optimized model has a prediction accuracy of up to 96 % on unknown strain-paths. Subsequently, the DNNSC model was implemented into the FE analysis through Fortran programming, and a benchmark simulation for thin sheet stamping was successfully performed. The simulation results demonstrated that the DNNSC model significantly improved prediction performance compared to conventional macroscale constitutive models. Especially, the ear height and plate thickness were accurately predicted with an accuracy of 91.85 % and 95.84 %, compared to only 68.85 % and 86.59 % achieved by the Yld model. Meanwhile, the simulation time was reduced to approximately one-tenth that of the fully coupled CP model, because the latter required calculating and homogenizing the mechanical responses of hundreds of grains at each integration point during the simulation. The DNNSC framework bridges the gap between CP models and FE simulations of plastic forming and breaks down the barrier between modeling and practical application. Furthermore, this framework can be extended to other materials by re-calibrating VPSC parameters and fine-tuning DNN parameters.}, archive = {J_EAAI}, author = {Ziwei Zhou and Liang Cheng and Huaidong Song and Haijing Guo and Ruolin Li and Lingyan Sun and Bin Tang}, doi = {10.1016/j.engappai.2025.112249}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112249}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multiscale constitutive modeling of anisotropic plasticity: Coupling the visco-plastic self-consistent model with the recurrent neural network and its implementation in finite element analysis}, volume = {161}, year = {2025}, }
• (2025). An online reinforcement learning method to improve control adaptability in robot-aided rehabilitation. EAAI, 161, 112248. (www)

  @article{ , abstract = {Rehabilitation robotics enables consistent and personalized therapy but still relies on complex, expert-driven tuning of control parameters. To address this, a reinforcement learning strategy based on Q-learning is proposed to autonomously adapt key parameters during upper-limb rehabilitation, without requiring prior task-specific knowledge. A systematic evaluation is conducted across combinations of control parameters (radial stiffness and execution time), performance-based reward functions (pointing accuracy and movement smoothness), and exploration strategies ( ɛ - greedy and Upper Confidence Bound). The Q-learning agent selects discrete actions (increase, decrease, or maintain the current value) for each control parameter, enabling real-time adaptation based on observed performance. The method is validated using a Kuka robotic arm in experiments involving 16 right-handed healthy subjects (13 males, 3 females) and 8 right-handed individuals simulating impaired motor behavior (5 males, 3 females). Motion signals are acquired through internal robot sensors, while a wearable physiological monitoring system define the Q-learning agent state. Reward improvement and exploration ratio are analyzed as key performance indicators and statistically compared across all tested conditions using the Mann–Whitney test. The results demonstrate that the proposed algorithm effectively adjusts control parameters online, with performance influenced by the reward function, exploration strategy, and selected control actions. Reward improvements of 0 . 11 ± 0 . 09 ( ɛ - greedy , reward based on pointing ability) and 0 . 13 ± 0 . 11 (reward based on smoothness, Upper Confidence Bound strategy) were observed in healthy subjects, indicating enhancements in pointing accuracy and movement smoothness. In simulated pathological cases, improvements of 0 . 08 ± 0 . 13 and 0 . 06 ± 0 . 16 were observed, respectively.}, archive = {J_EAAI}, author = {Rita Molle and Christian Tamantini and Clemente Lauretti and Emilio Maria Romano and Loredana Zollo}, doi = {10.1016/j.engappai.2025.112248}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112248}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An online reinforcement learning method to improve control adaptability in robot-aided rehabilitation}, volume = {161}, year = {2025}, }
• (2025). Fast prediction and compensation of curing deformation behaviours of composite parts with complex geometry based on neural operator on riemannian manifolds. EAAI, 161, 112247. (www)

  @article{ , abstract = {Controlling the curing deformation of composite parts is becoming increasingly challenging with the ever-increasing performance requirements of aerospace equipment. Mould surface compensation, which adjusts the mould surface to minimise the discrepancy between the cured part geometry and the nominal part geometry, has become a primary deformation control way in engineering. The existing mirror compensation methods focus on the deformation dominated by spring-in and are challenging for complex deformation modes. Surrogate model-based shape optimisation provides a feasible idea, but establishing a surrogate model to predict curing deformation fields on complex part geometries remains a challenge. Therefore, this study explores a novel neural operator-driven framework for fast curing deformation prediction and compensation. A clustering-based deformation field segmentation method is proposed to manipulate the mould surface morphing using limited design variables. The neural operator on Riemannian manifolds is introduced for the first time to establish the surrogate model between the mould surface and the curing deformation fields on complex part geometries. To control the global error distribution of the composite part, two error metrics are designed to optimise the mould surface by genetic algorithm. The verification results show that the proposed framework exhibits significant potential in predicting and compensating for the curing deformation field of composite parts with complex geometry.}, archive = {J_EAAI}, author = {Lu Chen and Yingguang Li and Jingyan Su and Weiwei Xu and Lin Hu and Gengxiang Chen and Xu Liu and Xiaozhong Hao}, doi = {10.1016/j.engappai.2025.112247}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112247}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fast prediction and compensation of curing deformation behaviours of composite parts with complex geometry based on neural operator on riemannian manifolds}, volume = {161}, year = {2025}, }
• (2025). Enhanced detection of acute leukemia: A hybrid machine learning framework with adaptive weight-optimized level set evolution. EAAI, 161, 112244. (www)

  @article{ , abstract = {In this research, a novel, multistage framework is developed for the automated detection of Acute Lymphoblastic Leukemia (ALL) and Acute Myeloid Leukemia (AML), addressing challenges due to the inclusion of overlapping cells, noise, and unwanted cells. It integrates an Adaptive Weight-Optimized Level Set Evolution (AWOLSE) scheme to ensures precise and accurate segmentation, coupled with a marker-controlled watershed algorithm to improve performance in regions of cell overlap and contact. It ensures more reliable cell boundary delineation. For classification, a hybrid model combining the strengths of Random Forest (RF) and Support Vector Machine (SVM) is employed that delivering superior performance by leveraging the complementary advantages of these classifiers. Furthermore, feature extraction with the Gray Level Co-occurrence Matrix (GLCM), followed by feature selection with Principal Component Analysis (PCA), aids in the identification of relevant features. The suggested methodology beats its competitors, providing the best ALL identification results on the Acute Lymphoblastic Leukemia Image Database (ALLIDB), with 99.07% accuracy, 97.96% recall, 100.00% specificity, and 100.00% precision. Similarly, on the American Society of Hematology (ASH) database, the technique achieves superior AML identification results with 96.25% accuracy, 95.00% recall, 97.50% specificity, and 97.44% precision.}, archive = {J_EAAI}, author = {Pradeep Kumar Das and Adyasha Sahu and Sukadev Meher and Rutuparna Panda and Ajith Abraham}, doi = {10.1016/j.engappai.2025.112244}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112244}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhanced detection of acute leukemia: A hybrid machine learning framework with adaptive weight-optimized level set evolution}, volume = {161}, year = {2025}, }
• (2025). Reinforcement learning-based secure tracking control for nonlinear interconnected systems: An event-triggered solution approach. EAAI, 161, 112243. (www)

  @article{ , abstract = {This paper proposes a secure tracking control method for nonlinear interconnected systems based on reinforcement learning, addressing both security constraints and mismatched conditions in such systems. By utilizing system augmentation techniques, the tracking control problem is reformulated into a stabilization problem for the augmented system, simplifying the original control task. Drawing inspiration from the Chinese philosophical principle, we further transform the problem into a zero-sum game framework. Moreover, a control barrier function (CBF) is incorporated into a cost function to ensure that system trajectories remain within a predefined safe region. An event-triggered mechanism is introduced, and an event-based safety Hamilton–Jacobi–Isaacs (HJI) equation is established. An adaptive single evaluation network is designed, leveraging experience replay techniques to solve the HJI equation. Finally, the Lyapunov method is employed to prove the uniform ultimate boundedness of both the tracking error and the weight error. The effectiveness of the proposed method is validated through numerical examples.}, archive = {J_EAAI}, author = {Chunbin Qin and Suyang Hou and Mingyu Pang and Zhongwei Wang and Dehua Zhang}, doi = {10.1016/j.engappai.2025.112243}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112243}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Reinforcement learning-based secure tracking control for nonlinear interconnected systems: An event-triggered solution approach}, volume = {161}, year = {2025}, }
• (2025). A quantized subtraction-convolution network for industrial lightweight edge interpretable diagnosis. EAAI, 161, 112241. (www)

  @article{ , abstract = {In addressing the prevalent challenges of delayed fault diagnosis, deployment complexity of advanced deep learning models on low-cost edge devices, and limited model interpretability, this study proposed an edge-cloud collaborative interpretable diagnosis based on quantized subtraction-convolution network. Firstly, an interpretable quantized subtraction-convolution network with lightweight three-layer structure is designed. Inspired by the adaptive spectral subtraction, a learnable sparse pulse kernel is designed to extract the fault feature as the subtraction layer. Subsequently, the convolution and classification layers are then integrated to produce interpretable results for efficient identification. To facilitate deployment and updates on edge devices, the quantized subtraction-convolution network is decomposed into a lightweight edge architecture and corresponding parameters. It can be deployed on edge devices, and an edge-cloud collaborative framework addresses its training and compression. Considering the sparse characteristics of quantized subtraction-convolution network, a sparse pulse quantization strategy and quantization-aware training technique were developed to compress the model parameters. Finally, a low-cost edge fault diagnosis node prototype with quantized subtraction-convolution network is designed for real-time edge fault diagnosis. Experiments shown that the proposed method achieved average accuracy of 99.88 percent with compression ratio of 9.5. The memory usage, floating-point operations per second, and average power consumption are respectively only 54 kilo binary byte, 0.053 mega binary byte, and 6.67 mJ. Actual gear edge diagnosis experiments confirmed the effectiveness, which can implement model inference in 0.045 s at the edge fault diagnosis node. It is anticipated that the proposed method will find extensive application in the industrial edge interpretable diagnosis.}, archive = {J_EAAI}, author = {Qihang Wu and Jun Luo and Wenbin Huang and Xiaoxi Ding}, doi = {10.1016/j.engappai.2025.112241}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112241}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A quantized subtraction-convolution network for industrial lightweight edge interpretable diagnosis}, volume = {161}, year = {2025}, }
• (2025). Lightweight multi-level feature integration transformer for image super-resolution. EAAI, 161, 112240. (www)

  @article{ , abstract = {Transformer-based methods have attracted significant attention in the field of image super-resolution. However, existing approaches typically concentrate on a single level of information for image reconstruction, and neglect the critical role of multi-level information in feature representation, resulting in the low utilization of the potential capabilities of Transformer. To address this limitation, we propose a novel Transformer model, named as Multi-Level Differential Window Attention Transformer (MLDAT), designed for multi-level information fusion. Specifically, our approach introduces a self-attention module based on differential windows to comprehensively extract and integrate feature information across varying window sizes. Additionally, we introduce a High-order Global Attention Module (HGAB) to combine the second-order attention with global self-attention, which facilitates the establishment of relationships between local features and the overall global feature context within an image while complementing local window information. Extensive experimental results demonstrate that our model can significantly improve the performance of image super-resolution , and achieves the better results compared with existing methods.}, archive = {J_EAAI}, author = {Shuheng Wang and Ziao Gong and Mengda Li and Shen Wu and Yilin He}, doi = {10.1016/j.engappai.2025.112240}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112240}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Lightweight multi-level feature integration transformer for image super-resolution}, volume = {161}, year = {2025}, }
• (2025). Three-dimensional object detection for autonomous driving via deep learning: A review. EAAI, 161, 112238. (www)

  @article{ , abstract = {With the rapid advancement of autonomous driving technology, there is an increasing demand for highly accurate and real-time vehicle perception systems. From an artificial intelligence (AI) perspective, three-dimensional (3D) object detection benefits from recent progress in deep neural networks, convolutional neural networks (CNNs), and transformer architectures, which provide powerful tools for feature extraction, spatial reasoning, and multi-modal data fusion. These AI techniques enable robust, uncertainty-aware predictions by effectively modeling complex sensor data. From an engineering standpoint, 3D object detectors serve as critical components in autonomous driving systems by translating AI-derived insights into real-time, high-accuracy vehicle perception. This paper reviews the research progress of deep learning-based 3D object detection algorithms in autonomous driving. First, commonly used data acquisition sensors are systematically categorized, and the most widely adopted 3D detection datasets and evaluation metrics are introduced; the standard 3D bounding-box representation and core network architectures are also explained. Second, algorithms are classified according to input data type: 1) single-modal methods, including vision-based, 3D data dimensionality reduction, point cloud-based, transformer-based, mamba-based, and hybrid point cloud approaches; 2) multi-modal methods, subdivided into serial fusion and parallel fusion strategies within network pipelines. The characteristics, contributions, and limitations of each category are summarized, and representative algorithms are compared across datasets to identify research trends. Finally, current challenges, such as sparse data handling, domain shifts, and real-time constraints, are examined, and prospective directions for future development are proposed.}, archive = {J_EAAI}, author = {Jiaqi Cai and Yiquan Wu}, doi = {10.1016/j.engappai.2025.112238}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112238}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Three-dimensional object detection for autonomous driving via deep learning: A review}, volume = {161}, year = {2025}, }
• (2025). Unscented kalman filter neural network with double-layer decomposition algorithm applied to the prediction of current efficiency in aluminum electrolysis processes. EAAI, 161, 112237. (www)

  @article{ , abstract = {Predictive modeling in predictive optimization control technology can effectively reduce energy consumption and improve production efficiency in the electrolytic aluminum process (EAP). Among the various modeling approaches, artificial neural networks (ANN) have been widely adopted in the EAP due to their strong capability to capture nonlinear relationships inherent in complex industrial systems. However, conventional ANN heavily rely on historical data to achieve optimal models, which limits their accuracy and generalization performance under strong disturbances and time-varying conditions. To address these problems, this article proposes a novel dynamic prediction model of unscented Kalman filter neural network with double-layer decomposition (UKFNN-DD), building upon the foundation of Kalman filter neural network (KFNN). First, singular value decomposition (SVD) is adopted to compute the square root of covariance matrices, enhancing the numerical robustness of the prediction algorithm and overcoming the shortcomings of traditional Cholesky decomposition. Furthermore, a dual-layer KFNN strategy is introduced to overcome the absence of one-step-ahead prediction in conventional state-space formulations. By applying a two-stage correction to the state variables using measurement data, the proposed method improves the adaptability of the model to external environmental variations. Finally, the prediction error of the state variables is optimized using a gradient descent algorithm, which improves the stability and reliability of the model’s prediction performance. Experimental results demonstrate that the proposed method significantly outperforms baseline methods, achieving a 4.38-fold reduction in mean absolute error (MAE) and a 77.29-fold reduction in the sum of squared errors (SSE), thereby verifying its superiority in dynamic prediction for aluminum electrolysis.}, archive = {J_EAAI}, author = {Xiaoyan Fang and Xihong Fei and Zhenyi Xu and Lei Su and Jing Wang}, doi = {10.1016/j.engappai.2025.112237}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112237}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unscented kalman filter neural network with double-layer decomposition algorithm applied to the prediction of current efficiency in aluminum electrolysis processes}, volume = {161}, year = {2025}, }
• (2025). Feature super-resolution-based method for small-scale target detection and segmentation in side-scan sonar. EAAI, 161, 112235. (www)

  @article{ , abstract = {The application of side-scan sonar in underwater target detection plays a significant role in marine engineering construction and ocean resource exploration. However, existing methods for small object detection often suffer from limited accuracy and robustness. To address this issue, we propose a feature super-resolution-based approach for detecting and segmenting small targets in side-scan sonar images. During network training, a feature super-resolution branch is first constructed. Both low-level and high-level features from the backbone of the You Only Look Once (YOLO) network are fed into this branch. After processing through an encoder-decoder architecture, a super-resolved feature map is reconstructed, and the network is optimized via backpropagation to enhance the ability to extract features of small targets. Furthermore, an exponential decay strategy is adopted to define the loss weights of different branches, establishing a branch-aware training mechanism to improve training effectiveness. During inference, the super-resolution branch is discarded to balance detection accuracy and inference efficiency. Experimental results demonstrate that the proposed method achieves superior performance in detecting and segmenting small-scale targets in side-scan sonar imagery, achieving state-of-the-art results on two public side-scan sonar small object datasets. Additionally, this approach can be extended as a training strategy for side-scan sonar target detection and segmentation networks. The source code is available at https://github.com/Yang-Code984/FSR_Sonar .}, archive = {J_EAAI}, author = {Zhiwei Yang and Jianhu Zhao and Xi Zhao and Chao Huang}, doi = {10.1016/j.engappai.2025.112235}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112235}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Feature super-resolution-based method for small-scale target detection and segmentation in side-scan sonar}, volume = {161}, year = {2025}, }
• (2025). Data-driven prediction of thermal and flow fields in magnetized boger-micropolar tri-hybrid nanofluids via deep artificial neural networks. EAAI, 161, 112232. (www)

  @article{ , abstract = {This study presents a deep learning framework for predicting the thermal and flow behavior of Boger micropolar tri-hybrid nanofluids under magnetized axial squeezing flow between two parallel disks. The fluid comprises gold, silver, and diamond nanoparticles dispersed in a water-based solution, forming a high-performance ternary hybrid heat transfer medium. A fully connected artificial neural network with three hidden layers (32–16–8 neurons) and hyperbolic tangent activation functions was trained on synthetic data generated using a boundary value problem solver. The model predicts four key physical quantities: axial velocity, streamwise velocity, microrotation, and temperature profile. Quantitative evaluation reveals excellent agreement between machine learning predictions and numerical benchmarks, with mean absolute errors consistently below 0.02 and coefficients of determination exceeding 0.99 across all outputs. Sensitivity analysis reveals the impact of the solvent fraction and vortex viscosity parameters on flow penetration and thermal stratification near the boundaries. This work introduces a novel combination of Boger–micropolar fluid dynamics, magnetohydrodynamic squeezing flow, and tri-hybrid nanoparticles into a unified ANN (Artificial Neural Networks) based surrogate, enabling accurate multi-output prediction of strongly coupled nonlinear fields. The proposed approach offers a scalable machine learning model for real-time optimization of nanofluid-based thermal management systems in engineering applications.}, archive = {J_EAAI}, author = {Mohammad Jalili and Hesam Ehsani and Ali Mirzagoli Ganji and Amirali Shateri and Mehdi mahboobtosi and Yuping Wu and Payam Jalili and Bahram Jalili and Davood Domiri Ganji}, doi = {10.1016/j.engappai.2025.112232}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112232}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Data-driven prediction of thermal and flow fields in magnetized boger-micropolar tri-hybrid nanofluids via deep artificial neural networks}, volume = {161}, year = {2025}, }
• (2025). Explainable knowledge recommendation for product innovation concept design based on knowledge graph and multi-task learning. EAAI, 161, 112231. (www)

  @article{ , abstract = {In the process of product innovation concept design, by enhancing the transparency of the knowledge recommendation process and providing designers with explainable recommendation results, decision-making time can be reduced, and design efficiency can be improved. Existing explainable knowledge recommendation methods can generate textual explanation information, but they often overlook the dynamic nature of the design process, which negatively affects the accuracy of the recommendations. To address this issue, the current study proposes a knowledge recommendation method for product innovation concept design based on knowledge graph (KG) and multi-task learning. Specifically, a TransD-based model is first constructed to perform the knowledge graph embedding (KGE) task. Then, KGE and Gated Recurrent Unit (GRU) are used to obtain dynamic knowledge demands across different temporal scales from the historical interactions of designers to enhance recommendation accuracy. A multi-task learning framework is subsequently introduced to enable feature sharing between the two tasks, improving the generalization and effectiveness of the model. Finally, an explanation strategy is designed by combining embedded knowledge and paths to provide optimal explainability information. Experimental results demonstrate that, compared with other state-of-the-art knowledge recommendation algorithms, the proposed method not only offers explainable recommendation results but also achieves higher accuracy, making it more suitable for real-world recommendation scenarios.}, archive = {J_EAAI}, author = {Yida Hong and Wenqiang Li and Hai Xiang and Chuanxiao Li and Changfu Wan}, doi = {10.1016/j.engappai.2025.112231}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112231}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Explainable knowledge recommendation for product innovation concept design based on knowledge graph and multi-task learning}, volume = {161}, year = {2025}, }
• (2025). Towards reproducible machine learning-based process monitoring and quality prediction research for additive manufacturing. EAAI, 161, 112223. (www)

  @article{ , abstract = {Additive manufacturing (AM) is increasingly adopted across industries for its ability to support design flexibility, rapid prototyping, and mass customization. Machine learning (ML)-based cyber-physical systems (CPSs) have been extensively developed to improve the print quality of AM. However, the reproducibility of these systems has not been thoroughly investigated due to a lack of formal evaluation methods. Reproducibility, a critical component of trustworthy artificial intelligence, is achieved when an independent team can replicate the findings or artifacts of a study using a different experimental setup and achieve comparable performance. In many publications, critical information necessary for reproduction is often missing due to a lack of comprehensive AM and ML domain knowledge, resulting in systems that fail to replicate the reported performance. Integrating AM and ML domain knowledge, this paper proposes a reproducibility investigation pipeline and a reproducibility checklist for ML-based AM process monitoring and quality prediction systems. Based on the CRoss Industry Standard Process (CRISP) methodology, the pipeline guides researchers through the key steps required to reproduce a study, while the checklist systematically extracts information relevant to reproducibility from the publication. We validated the proposed approach through two case studies: reproducing a fused filament fabrication warping detection system and a laser powder bed fusion melt pool area prediction model. Both case studies confirmed that the pipeline and checklist successfully identified missing information, improved reproducibility, and enhanced the performance of reproduced systems. Based on the proposed checklist and leveraging large language models, a reproducibility survey was conducted to assess the current reproducibility status within this research domain.}, archive = {J_EAAI}, author = {Jiarui Xie and Mutahar Safdar and Andrei Mircea and Bi Cheng Zhao and Yan Lu and Hyunwoong Ko and Zhuo Yang and Yaoyao Fiona Zhao}, doi = {10.1016/j.engappai.2025.112223}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112223}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Towards reproducible machine learning-based process monitoring and quality prediction research for additive manufacturing}, volume = {161}, year = {2025}, }
• (2025). Degradation-based predictive energy management for intelligent fuel cell hybrid electric vehicles with a novel deep reinforcement learning architecture. EAAI, 161, 112222. (www)

  @article{ , abstract = {The degradation of fuel cell systems (FCS) affects the energy management performance of fuel cell hybrid electric vehicles (FCHEVs), especially in the case of severe degradation. This study develops a novel predictive energy management architecture, which consists of the Extended Long Short-Term Memory (xLSTM) and Soft Actor-Critic (SAC). Specifically, the speed predictor is built using xLSTM network and leverages the speed and position information of the preceding and following vehicles. In order to provide a reliable basis for energy management in degradation scenarios, an FCS degradation model is developed, which enables the dynamic mapping of output efficiency curves under varying state-of-health (SOH) conditions. Sequentially, a SAC-based agent is employed as the energy management strategy (EMS), which innovatively takes full account of the impact of the FCS SOH on energy allocation, achieving proactive degradation-aware energy allocation. Simulation results demonstrate that the developed xLSTM architecture achieves a 71 % improvement in speed prediction accuracy compared to Transformer-based models within the Next Generation Simulation validated scenario. Moreover, at SOH = 90 %, the newly designed EMS reduces operational costs by 8.7 % and 10.7 % compared to conventional methods under the New European Driving Cycle and Worldwide Harmonized Light Vehicles Test Procedure, while decreasing FCS degradation costs by 17 % and 51 %, respectively. The innovative approach not only elevates energy efficiency but also prolongs FCS operational longevity via intelligent SOH-aware, which establishes a new paradigm for lifecycle-optimized energy management in FCHEVs.}, archive = {J_EAAI}, author = {Zhigen Nie and Zhuangfeng Shi and Yufeng Lian and Hao Song}, doi = {10.1016/j.engappai.2025.112222}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112222}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Degradation-based predictive energy management for intelligent fuel cell hybrid electric vehicles with a novel deep reinforcement learning architecture}, volume = {161}, year = {2025}, }
• (2025). Observer based fault tolerant control design for saturated nonlinear systems with full state constraints via a novel event-triggered mechanism. EAAI, 161, 112221. (www)

  @article{ , abstract = {Safety plays a crucial role in many promising applications as they must comply with stringent safety regulations while staying within physical limits. Therefore, safety fault-tolerant systems with saturated nonlinearities and input constraints are investigated in this paper. An event-triggered safety fault tolerant control (FTC) method based on adaptive dynamic programming (ADP) is proposed. The constrained state is modeled using a smooth mapping function, thus transforming the original system into an unconstrained framework. A fault observer is designed for unknown actuator faults occurring in the system. When an unknown fault occurs in the system, the actuator is compensated in real time according to the fault. The optimal safety value function is approximated by constructing a single critic neural network (NN). Then, a novel event-triggered mechanism is proposed, which allows the algorithm to obtain the optimal control law without constructing the event-triggered Hamilton–Jacobi–Bellman (HJB) equation. In addition, by adjusting the size of the parameter τ under the event triggered condition, different demands on the number of event triggers are realized, which in turn ensures the efficient use of resources while leading to the trade-off of optimal control. This paper also incorporates an empirical playback technique in the design of the critic’s update law to address the challenges associated with continuous incentive requirements. It is theoretically proven that all signals of the system are consistent with the limit bounds, thus ensuring the stability of the system.}, archive = {J_EAAI}, author = {Chunbin Qin and Mingyu Pang and Zhongwei Wang and Suyang Hou and Dehua Zhang}, doi = {10.1016/j.engappai.2025.112221}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112221}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Observer based fault tolerant control design for saturated nonlinear systems with full state constraints via a novel event-triggered mechanism}, volume = {161}, year = {2025}, }
• (2025). Cybersecurity enhancement using conditional generative adversarial network with transformer-based conditional variational autoencoder. EAAI, 161, 112220. (www)

  @article{ , abstract = {Since, Artificial Intelligence is highly developing and concatenating into several domains, cybersecurity is an important field of delivering both the advantages and disadvantages. In addition to this, Artificial Intelligence is applied in a wide variety of applications like healthcare sector, content creation and entertainment and financial industries. Therefore, this work finds the efficiency of Artificial Intelligence -oriented cybersecurity metrics in succeeding the digital environment over elevating cyber threats. Here, the developed models consist of two different stages while implementing the model. In the first stage, the essential dataset is assembled from the benchmark data source. These datasets are assembled by using Generative Artificial Intelligence (Gen Artificial Intelligence networks). Consequently, the raw data is given as an input to Conditional Hybrid Network for cybersecurity enhancement. Further, the Transformer-based Conditional Variational Autoencoders with Spatial-temporal Attention are designed for feature extraction that is subjected to the Conditional Generative Adversarial Network for classifying the cyber attacks. Henceforth, the developed network is evaluated and designed with multiple measures. Comparing baseline models, the suggested network obtains higher performance for developing security over cyber networks.}, archive = {J_EAAI}, author = {Prithvipal Singh and Sandeep Singh and Gurupdesh Singh and Amritpal Singh}, doi = {10.1016/j.engappai.2025.112220}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112220}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Cybersecurity enhancement using conditional generative adversarial network with transformer-based conditional variational autoencoder}, volume = {161}, year = {2025}, }
• (2025). A modal transfer enhanced deep learning for structural dynamic response with sparse spatial data. EAAI, 161, 112218. (www)

  @article{ , abstract = {Sparse spatial data in structural health monitoring (SHM) presents significant challenges due to large spatial gaps and high uncertainty, which hinder the generalization ability and performance of traditional data-driven methods that rely on dense sensor networks. To address this issue, this paper proposes a novel modal transfer-enhanced deep learning (MT-DL) model for the reconstruction and prediction of structural dynamic responses. The core idea is to integrate physical modal information (mode shapes) into the deep learning framework via transfer learning. This integration enables the model to capture the spatial correlation among structural nodes, even under extremely limited data conditions. The MT-DL model is initially validated on a simply supported beam with varying stiffness, where it demonstrates significantly higher reconstruction accuracy compared to both conventional deep learning (DL) approaches and the graph convolutional network (GCN) model under sparse sensor conditions. To demonstrate the model's robustness and generalizability, it is further applied to the dynamic response reconstruction of a continuous beam subjected to moving loads, a plate with complex boundary conditions under impact loading, and a slender flexible riser undergoing vortex-induced vibration (VIV). The results show that the proposed MT-DL model, by incorporating physical principles into data-driven learning, not only enhances prediction accuracy with minimal data but also offers improved physical interpretability. This approach provides a promising solution for structural monitoring in scenarios where dense sensor instrumentation is impractical or cost-prohibitive.}, archive = {J_EAAI}, author = {Yangyang Liao and Yajuan Xie and Hesheng Tang and Zihan Xia and Songtao Xue}, doi = {10.1016/j.engappai.2025.112218}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112218}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A modal transfer enhanced deep learning for structural dynamic response with sparse spatial data}, volume = {161}, year = {2025}, }
• (2025). A novel deep learning pipeline for coronary artery analysis in X-ray angiography. EAAI, 161, 112217. (www)

  @article{ , abstract = {X- ray angiography is a primary and essential assistive method for the accurate diagnosis of coronary blockage. Segmentation of coronary arteries and analysis of coronary artery blockage are the two primary tasks performed by cardiologists. We have proposed Coronary Artery Segmentation, Blockage Detection, and Measurement (CASBloDaM) a state-of-the-art pipeline integrated deep learning framework with conventional image processing techniques to achieve accurate coronary artery analysis in X-ray angiographic images. A new private dataset of 214 X-ray angiographic images is prepared for training, validation, and testing of the model by manual pixel annotation. The UNet3+ deep learning model is trained on X-ray angiographic images for accurate artery segmentation while multiple image processing techniques are developed to perform blockage detection and measurement. The model shows excellent performance in artery segmentation and the dice score of 0.989, accuracy of 0.985, specificity of 0.913 and sensitivity of 0.847 is achieved during the testing which is superior when compared against the previous reported works. An accuracy of 0.853, 0.882 and 0.725 are achieved by the proposed catheter detection, sub-artery removal and blockage detection algorithms respectively when evaluated for 150 test images. The proposed methodology achieved a Mean Square Error (MSE) of 28.66 and an R-square of 0.81 for blockage percentage estimation, and an MSE of 69.91 with an R-square of 0.99 for blockage location, compared to manual calculations. The novelty of the present work is the end-to-end integrated framework which can accurately perform coronary artery analysis in X-ray angiographic images and propose its clinical usage.}, archive = {J_EAAI}, author = {Karan V. Padariya and Abhishek Raval and Pranay Soni and Harsh Kapadia}, doi = {10.1016/j.engappai.2025.112217}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112217}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel deep learning pipeline for coronary artery analysis in X-ray angiography}, volume = {161}, year = {2025}, }
• (2025). A multi-teacher knowledge distillation-based framework for long-term respiratory monitoring and prediction with a novel flexible wearable sensor in healthcare engineering. EAAI, 161, 112216. (www)

  @article{ , abstract = {Respiratory monitoring plays a critical role in early health warnings and preventive care, offering significant potential for advancements in healthcare engineering. Wearable respiratory monitoring devices, known for their compact design, portability, and real-time capabilities, face challenges such as limited long-term comfort, environmental interference, and signal inaccuracies. In this study, we propose a novel wearable respiratory monitoring framework, RAMP, which integrates an innovative artificial muscle-based flexible sensor system with advanced deep learning modules. The system is designed to reconstruct and analyze respiratory data across various human activities and predict long-term respiratory function. Utilizing a multi-teacher knowledge distillation mechanism, the framework optimizes a student model for enhanced prediction accuracy. Experimental results demonstrate the mean absolute percentage error (MAPE) of 7.28 and mean absolute error (MAE) of 11.92, highlighting the feasibility and effectiveness of system. This work advances the development of portable health monitoring devices and provides a robust foundation for long-term respiratory activity assessment and forecasting, contributing to the broader field of healthcare engineering and personalized medicine.}, archive = {J_EAAI}, author = {Ke Li and Qing Wang and Haoke Liu and Mingke Wang and Suiyuan Zhu and Xiang Wang and Jing Qin}, doi = {10.1016/j.engappai.2025.112216}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112216}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-teacher knowledge distillation-based framework for long-term respiratory monitoring and prediction with a novel flexible wearable sensor in healthcare engineering}, volume = {161}, year = {2025}, }
• (2025). Dynamic mask network based on spiking neural convolutional model for missing modality brain tumor segmentation. EAAI, 161, 112215. (www)

  @article{ , abstract = {Brain tumor segmentation is a medical image processing task aimed at accurately locating and isolating tumor regions from brain scan images (e.g., Magnetic Resonance Imaging, MRI) in order to help doctors in diagnosis, treatment planning and surgical navigation. Automatic brain tumor segmentation is extremely challenging due to incomplete feature representation in the case of missing modalities and insufficient inter-modal information interaction. To this end, this paper proposes a novel dynamic threshold mask Transformer network for the missing modality brain tumor segmentation task, which is designed based on the nonlinear spiking neural convolutional model. The network consists of four independent encoders and a shared decoder to extract the features of each modality and perform shared representation learning. Among them, the dynamic threshold mask Transformer introduces learnable embedding vectors, generates dynamic masks on top of static masks to achieve fine-grained feature filtering, and enhances the ability of inter-modal information interaction. The adaptive gating weighting module and the channel cross spiking neural P attention module fuse modal features layer by layer in both spatial and channel dimensions to strengthen the modeling capability of local and global features. We conducted extensive comparative experiments on different missing modal cases in the BraTS2020 and BraTS2018 datasets. The experimental results show that the method effectively improves the robustness of missing modalities and the performance of brain tumor segmentation while maintaining the computational efficiency, and has good generalization ability and practicality.}, archive = {J_EAAI}, author = {Junjie Li and Rui Cai and Bing Li and Hong Peng}, doi = {10.1016/j.engappai.2025.112215}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112215}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Dynamic mask network based on spiking neural convolutional model for missing modality brain tumor segmentation}, volume = {161}, year = {2025}, }
• (2025). Enhancing accuracy in fall detection and prediction for elderly individuals using ensemble wavelet neural network and maximal overlap discrete wavelet transform. EAAI, 161, 112213. (www)

  @article{ , abstract = {Falls are the second-largest risk factor for the health of the elderly. Various researchers have proposed approaches for monitoring health and falls using only the Internet of Things (IoT) and sensors. However, relying on a single sensor limits accuracy in fall risk prediction. Data from multiple sensors can be used to classify human posture through learning-based algorithms stemming from machine learning. Recently, posture-detecting sensors have become increasingly popular, with InvenSense's Inertial Measurement Unit 9250 (IMU 9250) being a notable example for detecting elderly motion and transmitting data to a computer. This paper proposes the Ensemble Wavelet Network Framework (EWNNET) for fall prediction and detection using the Maximal Overlap Discrete Wavelet Transform (MODWT). The inertial measurement unit (IMU) dataset, containing real-time posture data from seven sensors, is split into 80 % for training and 20 % for testing. The EWNNET approach is compared with other machine-learning methods for fall detection, showing improved accuracy, sensitivity, precision, F-score, and specificity, with a 96.7 % classification success rate.}, archive = {J_EAAI}, author = {Safa Hussein Mohammed and Yangyu Fan and Guoyun Lv and Shiya Liu}, doi = {10.1016/j.engappai.2025.112213}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112213}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing accuracy in fall detection and prediction for elderly individuals using ensemble wavelet neural network and maximal overlap discrete wavelet transform}, volume = {161}, year = {2025}, }
• (2025). Leakage localization methodology based on dynamic pressure signal for subsea pipeline. EAAI, 161, 112212. (www)

  @article{ , abstract = {Leakage is one of the most critical failure forms of subsea pipeline. The accurate leakage localization is of great significance to ensure the safe and reliable transportation of subsea pipeline. Leakage locations are considered to be discretely distributed along the subsea pipeline. However, an overabundance of nodes in leakage localization model is caused by excessive discretization. The performance of a leakage localization model with much localization points is poor. Furthermore, the data collected in site usually contains a lot of noise which reduce the effectiveness of leakage localization. A leakage localization methodology based on dynamic pressure signal for subsea pipeline is proposed in this paper. A noise reduction model based on variational mode decomposition (VMD) algorithm combined with power spectral density (PSD) is established to reduce noise of leakage signal. An improved K-means grouping model is developed to mining data for inherent similarity and cluster leakage characteristic. It improves robustness of the leakage localization model. A radial basis function (RBF) neural network leakage localization model optimized by the pelican optimization algorithm (POA) is used to identify the location of leakage. A leakage experiment is used to study performance of this methodology. The localization accuracy of the proposed leakage localization methodology is more than 90 %, the localization error is less than 16 cm. After neural network combined with improved grouping model, average leakage localization accuracy increased by 15.65 %, average absolute error decreased 8.64 cm. The proposed methodology provides an effective tool for leakage localization of critical equipment in subsea production system.}, archive = {J_EAAI}, author = {Guowei Ji and Baoping Cai and Xuelin Liu and Yi Jiang and Yixin Zhao and Qingping Li and Lei Gao and Kaizheng Wu}, doi = {10.1016/j.engappai.2025.112212}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112212}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Leakage localization methodology based on dynamic pressure signal for subsea pipeline}, volume = {161}, year = {2025}, }
• (2025). Enhancing entity and relation extraction with dynamic hard negative augmentation framework. EAAI, 161, 112211. (www)

  @article{ , abstract = {Entity and relation extraction, a fundamental task in information extraction, plays a crucial role in modeling unstructured text by identifying meaningful entities and their semantic relationships. While existing methods have shown effectiveness, they still face challenges in accurately identifying entity boundaries and extracting complex relationships. These challenges primarily arise from current contrastive learning approaches, which uniformly handle all negative samples in boundary detection and relation extraction without emphasizing the learning of hard negative samples. Additionally, the scarcity of hard negative samples limits the exploration of the state space near the anchors. To tackle these challenges, we introduce Dynamic Hard Negative Augmentation, an innovative framework designed to strategically explore and generate hard negative samples, thereby enhancing the learning of challenging cases through adaptive contrastive learning. During the negative sample augmentation process, we employ adversarial training to explore underrepresented areas of hard negative samples, generating a comprehensive coverage of the hard negative sample space to effectively explore the state space. We further introduce a dynamic enhancement mechanism that continuously optimizes the proportion of hard negative samples during training, ensuring targeted learning of these challenging cases.}, archive = {J_EAAI}, author = {Qibin Li and Shengyuan Bai and Nai Zhou and Nianmin Yao}, doi = {10.1016/j.engappai.2025.112211}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112211}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing entity and relation extraction with dynamic hard negative augmentation framework}, volume = {161}, year = {2025}, }
• (2025). NashDQNSleep: Nash-based deep Q-network adaptive sleep scheduling for energy efficiency and age of information optimization in industrial internet of things. EAAI, 161, 112210. (www)

  @article{ , abstract = {Fresh and timely data is essential for the sustainable operation of Industrial Internet of Things (IIoT) systems, which support real-time monitoring and decision-making tasks. Sensor nodes typically need to remain active for long durations to maintain data freshness, resulting in high energy consumption. Traditional sleep scheduling methods often trade off energy savings with data freshness, and many rely on centralized sink-node coordination, which can create communication bottlenecks and increase energy use. This paper introduces NashDQNSleep , a novel decentralized sleep scheduling algorithm that uniquely combines game-theoretic modeling with deep reinforcement learning to address these challenges. Unlike conventional DQN approaches that optimize node behavior independently or centrally, NashDQNSleep formulates the scheduling problem as a general-sum stochastic game and uses Deep Q-Networks (DQNs) to compute Nash equilibria. This enables sensor nodes to make autonomous, yet strategically coordinated decisions based on peer-to-peer information exchange, achieving a stable and efficient balance between energy consumption and data freshness without centralized control. Extensive simulations on diverse IIoT network topologies demonstrate that NashDQNSleep improves energy efficiency by up to 35%, reduces Age of Information (AoI) by 40%, and increases successful data transmissions by 15%–20% compared to state-of-the-art decentralized and centralized methods. These results establish NashDQNSleep as an effective, scalable, and practical solution for reliable and sustainable IIoT operations.}, archive = {J_EAAI}, author = {Partha Sarathi Banerjee and Soumyapriya Goswami and Debashis De}, doi = {10.1016/j.engappai.2025.112210}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112210}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {NashDQNSleep: Nash-based deep Q-network adaptive sleep scheduling for energy efficiency and age of information optimization in industrial internet of things}, volume = {161}, year = {2025}, }
• (2025). A novel contrastive learning framework for multi-parameter optimization in 3D printing. EAAI, 161, 112209. (www)

  @article{ , abstract = {Additive manufacturing (3D printing) revolutionizes prototyping and production through unparalleled material efficiency. However, part quality remains highly sensitive to parameter variations, where subtle deviations induce defects, material waste, and process instability. Traditional quality control methods – reliant on rule-based heuristics or manual inspections – fail to address complex multi-parameter interactions and fine-grained anomalies, limiting industrial scalability. This article focuses on an important issue in 3D printing: How can we develop a robust, automated framework to simultaneously detect and classify subtle multi-parameter anomalies in 3D printing, overcoming the limitations of manual and single-defect-focused approaches? We propose a supervised contrastive learning framework integrating Vision Transformers (ViT) to learn discriminative feature representations for multi-parameter optimization. By maximizing intra-class similarity and inter-class separation, our model captures nuanced variations across printing scenarios. The ViT architecture processes real-time printing images, while contrastive loss ensures compact feature clusters for “Low”, “Optimal”, and “High” parameter classes. Experimental evaluations on open-source datasets demonstrate our framework achieves 8.45% accuracy, outperforming conventional CNNs by 10.11%. Real-world validation shows robust performance across critical parameters: flow rate (86.5% accuracy in nominal ranges), feed rate (87% accuracy), and extrusion temperature (90% accuracy at optimal settings). The ViT’s self-attention mechanism enables precise detection of localized anomalies, such as under-extrusion and layer misalignment.}, archive = {J_EAAI}, author = {Jieyang Peng and Simon Kreuzwieser and Dongkun Wang and Andreas Kimmig and Zhi Fan and Jianing Li and Jivka Ovtcharova}, doi = {10.1016/j.engappai.2025.112209}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112209}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel contrastive learning framework for multi-parameter optimization in 3D printing}, volume = {161}, year = {2025}, }
• (2025). A lightweight model based on multi-scale feature fusion for ultrasonic welding surface defect detection. EAAI, 161, 112208. (www)

  @article{ , abstract = {Ultrasonic welding technology is crucial in industrial and medical fields, relying on precise surface defect detection for quality assurance. Traditional methods suffer from low accuracy, efficiency, high costs, and complex implementation. Additionally, current neural networks for ultrasonic surface defect detection struggle to balance parameter optimization with detection accuracy. To solve this problem, we proposed a lightweight model based on multi-scale feature fusion for the Ultrasonic Weld Surface Defect Detection Network (UWSDNet). First, the feature extraction module with reparameterization technology (FRT) and application of efficient multi-scale attention (EMA) are proposed to alleviate network redundant parameters and computational overhead brought by welding background. Secondly, the multi-core feature enhancement module (MCM) is introduced. It enhances multi-scale object detection with fewer parameters to cope with the actual edge deployment of ultrasonic welding. Finally, the lightweight asymmetric detection head (LADH) and contextual and spatial feature calibration network (CSFCN) are introduced into the network. To improve the multi-core dimensional feature capture capability, to solve the problem of large size span of ultrasonic welding surface defects. Experimental evaluations on a self-built ultrasonic welding wire harness defect dataset show that UWSDNet achieves the mean average precision (mAP) of 88.9%, the precision of 95.6% with parameters of 12.7M. In addition, UWSDNet achieves excellent performance on the publicly available NEU-DET dataset, demonstrating strong generalization and application potential in industrial defect detection.}, archive = {J_EAAI}, author = {Rui Liu and Lun Zhao and Yu Ren and Zhonghua Shen and Liya Li and Jianfeng Luo and Zeshan Abbas}, doi = {10.1016/j.engappai.2025.112208}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112208}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight model based on multi-scale feature fusion for ultrasonic welding surface defect detection}, volume = {161}, year = {2025}, }
• (2025). Artificial neural network solution for magnetohydrodynamic ternary hybrid second-grade nanofluid in rotating jeffery-hamel flow under heat generation: Entropy generation analyses. EAAI, 161, 112207. (www)

  @article{ , abstract = {This research-work examines entropy generation in Magnetohydrodynamic (MHD) ternary hybrid nanofluid under rotating Jeffery-Hamel flow with heat generation alongside thermal radiation. Using appropriate similar transformation equations, the original partial differential equations (PDEs) are converted into nonlinear ordinary differential equations (ODEs), while Artificial Neural Networks (ANNs), which highly extensively used as universal function approximators, is considered in this investigation. In fact, to provide efficient solutions for the nonlinear transformed problem of the magnetohydrodynamic ternary hybrid second-grade nanofluid in rotating Jeffery-Hamel flow, solvers consider log-sigmoid, radial basis and tan-sigmoid activation functions, optimized with an interior point method to generate optimal weights of each considered ANN model. The model relies on assumptions of incompressible steady flow together with negligible viscous energy loss and usage of Rosseland radiation approximation. The predictions made through the Artificial Neural Network (ANN) match numerical benchmarks successfully thus demonstrating better accuracy when modeling velocity, temperature and entropy profiles. The research data indicates that systems attaining better thermal qualities operate effectively under elevated nanoparticle levels and radiation conditions which provides fundamental knowledge to enhance thermal system design. The results from the proposed schemes match numerical solutions precisely thus showing their high accuracy for evaluating ternary hybrid nanofluid flow and heat behavior. Results show that the velocity and temperature profiles of the Magnetohydrodynamic (MHD) ternary hybrid nanofluid reveal significant improvements at higher thermal volume fractions of nanoparticles. The rate of flow increases, perhaps because inertial forces prevail over viscous forces, thereby improving flow dynamics. With higher Radiation numbers associated with higher permeability, the local temperature gradients decrease promoting heat transfer rates within the fluid.}, archive = {J_EAAI}, author = {Nouar Ahcene and Amar Dib and Farhan Lafta Rashid and Kezzar Mohamed and Mohamed Rafik Sari and Manal Elzain and Hamiden Abd El-Wahed Khalifa}, doi = {10.1016/j.engappai.2025.112207}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112207}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Artificial neural network solution for magnetohydrodynamic ternary hybrid second-grade nanofluid in rotating jeffery-hamel flow under heat generation: Entropy generation analyses}, volume = {161}, year = {2025}, }
• (2025). Intelligent compound fault decoupling of rolling bearing based on parallel capsule network. EAAI, 161, 112206. (www)

  @article{ , abstract = {Due to the complex working environment of rolling bearings, various components such as inner ring, outer ring, rolling element and cage in bearing may interact with each other, leading to a compound fault formed by a variety of single fault coupling. Most methods generally regard compound faults as single faults, ignoring the interaction between single faults, which is not conducive to decoupling compound faults into multiple single faults and formulating maintenance plans. Moreover, the capsule network requires stacking multiple capsule layers to enhance performance, which significantly increases model parameters and consumes substantial memory resources. Therefore, a compound fault intelligent decoupling method based on parallel capsule network combining dynamic routing and attention routing is proposed in this study. Firstly, the Omni-Scale block is added to the feature extraction part, which can cover different sizes of receptive fields to enhance the feature extraction ability of the network. Secondly, an attention routing module is proposed, which realize the transmission of information from low-level capsules to high-level capsules by calculating the correlation between the same layers. Finally, the parallel capsule decoupling layer is constructed by using dynamic routing and attention routing. This method is especially suitable for practical engineering scenarios where compound bearing fault samples are limited and computational resources are constrained, providing a lightweight and effective solution for intelligent fault diagnosis. Experimental results show that the proposed method significantly reduces model complexity while maintaining high diagnostic performance under small-sample conditions, with the ablation study further confirming the meaningful contribution of each core module.}, archive = {J_EAAI}, author = {Renwang Song and Chenyu Jiao and Hui Shi and Linying Chen}, doi = {10.1016/j.engappai.2025.112206}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112206}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Intelligent compound fault decoupling of rolling bearing based on parallel capsule network}, volume = {161}, year = {2025}, }
• (2025). Deep reinforcement learning-based strategic bidding in electricity markets via variational autoencoder-assisted competitor behavior learning. EAAI, 161, 112205. (www)

  @article{ , abstract = {In a deregulated electricity market, self-interested producers have incentives to offer strategically for maximizing their own profits. While deep reinforcement learning (DRL) has shown great potential for solving such strategic bidding problems, existing methods typically oversimplify strategic action spaces and neglect the influence of competitors' offering behaviors. To bridge these gaps, this paper proposes a novel DRL-based framework to model and solve the strategic bidding problem of an individual producer by jointly considering price-quantity offering actions and the dynamic behaviors of market competitors. First, a bilevel optimization model is formulated to incorporate offering actions on price-quantity pairs. Then, a data-driven framework that combines a variational autoencoder with a density-based clustering method is proposed to learn and capture competitors' offering behaviors. Finally, an imitation learning-integrated DRL algorithm is developed to improve learning stability and solution quality for strategic bidding with price-quantity actions and competitors' offering behaviors. Case studies on the IEEE-30 bus system show that the proposed framework obtains a 28.12 k$ (24.25 %) increase in average profit compared to the existing approach, demonstrating its effectiveness and adaptability under dynamic market conditions.}, archive = {J_EAAI}, author = {Fei Hu and Yong Zhao and Yaowen Yu and Yuanzheng Li}, doi = {10.1016/j.engappai.2025.112205}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112205}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep reinforcement learning-based strategic bidding in electricity markets via variational autoencoder-assisted competitor behavior learning}, volume = {161}, year = {2025}, }
• (2025). A deep-supervised framework utilizing edge features and global texture features for distortion rectification in fisheye images. EAAI, 161, 112204. (www)

  @article{ , abstract = {Fisheye cameras, with their ultra-wide field of view, are increasingly deployed in urban environments. However, the severe radial distortion inherent in fisheye imagery presents formidable challenges for downstream computer vision tasks, including object recognition, motion estimation, and semantic segmentation. To address this issue, a deeply supervised framework for distortion rectification that leverages edge features and global texture information (DR-EGT) is proposed, which uniquely integrates edge-aware structural features with global texture information under a unified deep supervision strategy. Unlike conventional approaches that rely solely on texture priors or geometric assumptions, DR-EGT introduces a hierarchical supervision mechanism that simultaneously leverages low-level edge contours and high-level texture semantics to guide the distortion correction process. This joint supervision enables the network to learn the distortion flow field of the entire image, which facilitates the reconstruction of geometrically accurate and perceptually sharp undistorted images. On the Places2 dataset, DR-EGT outperforms existing advanced methods, achieving a 12.12 % increase in PSNR (Peak Signal-to-Noise Ratio), 7.44 % improvement in SSIM (Structural Similarity Index Measure), 16.42 % reduction in MAE (Mean Absolute Error), and a 78.17 % gain in FID (Fréchet Inception Distance), demonstrating its superior reconstruction fidelity and perceptual quality. The results demonstrate the ability of DR-EGT not only correct complex fisheye distortion but also suppress post-correction artifacts and visual degradation.}, archive = {J_EAAI}, author = {Yicheng Chen and Shuang Li and Yuhuan Gu and Chang Feng and Changhai Zhai}, doi = {10.1016/j.engappai.2025.112204}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112204}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A deep-supervised framework utilizing edge features and global texture features for distortion rectification in fisheye images}, volume = {161}, year = {2025}, }
• (2025). Spectrum prior-based and visibility fusion method for underwater image enhancement. EAAI, 161, 112203. (www)

  @article{ , abstract = {When light propagates in water, it undergoes scattering and absorption phenomena, which typically result in haze, high blur, low contrast and color distortion, making it extremely challenging to obtain high-quality images. To address these issues, many existing methods target image enhancement by correcting specific aspects such as color shift or contrast. However, challenges like poor visibility and low-light conditions are often overlooked. In this paper, we proposed a spectrum prior-based and visibility fusion method (SPV) to enhance underwater images in terms of color, contrast, and visibility. Unlike existing methods, SPV complements the advantages of both physical and non-physical models, comprehensively addressing the problems of reduced visual visibility, color distortion, and low contrast caused by low-light environments, thereby significantly improving the overall image quality. We proposed a dehazing module based on spectral information priors, which reliably restores image quality under complex water conditions. Additionally, we introduced a color correction module based on human color perception and employed morphological operations, effectively solving the issues of color shift and unclear contours in underwater images. Furthermore, we proposed a visibility enhancement module based on the fuzzy c-means clustering method to improve image contrast and visibility, particularly under low-light conditions. Finally, through a detail enhancement fusion module, we simultaneously addressed problems related to color shift, low contrast, and low visibility. SPV showed excellent performance in application tests including feature point matching, geometric rotation estimation, and edge detection. Comparative experiments on four real underwater datasets against 14 advanced enhancement methods demonstrated promising results.}, archive = {J_EAAI}, author = {Qifeng Liu and Xin Yan and Lu Shen and Qiang Li}, doi = {10.1016/j.engappai.2025.112203}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112203}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Spectrum prior-based and visibility fusion method for underwater image enhancement}, volume = {161}, year = {2025}, }
• (2025). An integrated multi-criteria decision-making approach with unknown weight information under probabilistic dual-hesitant pythagorean fuzzy environment and its application to supplier selection. EAAI, 161, 112202. (www)

  @article{ , abstract = {The preference ranking organization method for enrichment evaluation (PROMETHEE) has been proved to be one of the most effective techniques to rank alternatives of multi-criteria decision-making (MCDM) problems. However, the existing PROMETHEE cannot accurately adjust the representation range of uncertain information. Besides, the weight determination in PROMETHEE heavily relies on the decision matrix of alternatives on the criteria. Moreover, the information aggregation in PROMETHEE models lacks consideration of the interrelationship among criteria. To address the aforementioned shortcomings, this paper introduces a novel MCDM method that integrates the best-worst method (BWM) and PROMETHEE to help the decision-maker (DM) select the optimum alternative under the probabilistic dual-hesitant Pythagorean fuzzy (PDHPF) environment. Firstly, we extend PROMETHEE method to PDHPF scenario, which not only helps the DM depict subjective evaluations, but also provides the DM with a laxer constraint to present decision information. Secondly, the PDHPF power weighted Hamy mean operator (PDHPFPWHM) is utilized to aggregate the preference information of PROMETHEE. Additionally, the BWM method is utilized and extended to the PDHPF environment for acquiring optimal weights of criteria. The high efficiency and consistency of BWM are well-suited for the complex PDHPF environment. Finally, the method is applied to a semiconductor supplier selection case to demonstrate the validity, superiority, and feasibility of the proposed approach.}, archive = {J_EAAI}, author = {Huzhi Xue and Haihua Xie and Butian Zhao and Jun Wang}, doi = {10.1016/j.engappai.2025.112202}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112202}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An integrated multi-criteria decision-making approach with unknown weight information under probabilistic dual-hesitant pythagorean fuzzy environment and its application to supplier selection}, volume = {161}, year = {2025}, }
• (2025). A novel approach for understanding the parking demand for internal access roads in hub car parks. EAAI, 161, 112201. (www)

  @article{ , abstract = {The hubs serve as pivotal nodes within urban transport networks, and the car parks associated with these hubs constitute an integral component. The prevalent practice in China of utilizing internal access roads within car parks for passenger pick-up by online car-hailing vehicles has engendered a novel category of parking demand. However, this development presents a significant challenge to the efficient operation of hub car parks. A thorough analysis and comprehension of the parking demand for internal access roads is essential in devising suitable management strategies to enhance parking efficiency. Initially, this study categorizes the parking demand for internal access roads into three distinct groups, applying parking duration and the number of entries as classification criteria. Subsequently, considering the number of train frequencies, the categorized parking demands are forecasted independently utilizing the Long Short-Term Memory (LSTM) model, complemented by the Shapley Additive exPlanations (SHAP) method for model interpretation. Ultimately, Vector Autoregression (VAR) is employed to investigate the interaction mechanism between the parking demand for parking spaces and the parking demand for internal access roads. The findings indicate that optimal predictive performance is attained when employing a time interval of 15 min and an input step size of 4. Competition between the parking demand for internal access roads and the parking demand for parking spaces primarily occurs within respective categories, underscoring the need for external interventions when these demands are imbalanced.}, archive = {J_EAAI}, author = {Qianyi Hu and Weidong Liu and Chenyu Yan and Chu Zhang and Jun Chen and Changyin Dong}, doi = {10.1016/j.engappai.2025.112201}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112201}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel approach for understanding the parking demand for internal access roads in hub car parks}, volume = {161}, year = {2025}, }
• (2025). Machine learning for disease-specific prediction of high-cost patients. EAAI, 161, 112200. (www)

  @article{ , abstract = {High-cost patients incur disproportionately high medical expenses, and identifying them proactively is crucial for effective healthcare management. While previous research has focused on identifying high-cost patients based on overall expenditure, there has been a lack of studies analyzing them in the context of specific diseases. This study addressed this gap by leveraging data from the National Health Insurance Service (NHIS) of South Korea, spanning 2015 to 2019, to develop predictive models for identifying these patients. We trained models using data from 880,000 individuals to predict high-cost patients in 2019 using resource-efficient machine learning algorithms such as Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Neural Networks (NN) that minimize computational overhead, with undersampling techniques applied to handle data imbalance. We focused on the six major disease categories that account for the highest medical expenditures in South Korea: diseases of the musculoskeletal system (DMS), circulatory system (DCS), eye and ear (DEA-DEM), digestive system (DDS), genitourinary system (DGS), and respiratory system (DRS). We discovered that disease-specific analyses revealed important predictive factors that were not apparent in aggregate analyses. For example, hemoglobin levels emerged as crucial predictors for DCS, while body mass index (BMI) proved essential for DMS prediction. These findings enhance our understanding of the factors contributing to high medical costs and provide a foundational framework for healthcare providers and policymakers to develop more targeted and effective health management strategies.}, archive = {J_EAAI}, author = {Inwoo Tae and Hyeongwoo Kong and Junghye Lee and Yongjae Lee}, doi = {10.1016/j.engappai.2025.112200}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112200}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Machine learning for disease-specific prediction of high-cost patients}, volume = {161}, year = {2025}, }
• (2025). Optimizing mean conditional value-at-risk portfolios through deep neural network stock prediction. EAAI, 161, 112198. (www)

  @article{ , abstract = {Portfolio optimization is essential in financial decision-making, requiring a balance between risk minimization and return maximization. Effective stock selection significantly influences portfolio performance. Traditional methods often struggle to effectively integrate advanced risk assessment techniques with stock selection. To enhance portfolio diversification and improve risk-adjusted returns, this study integrates deep learning-based stock selection with the mean conditional Value-at-Risk (MCVaR) model and entropy constraints to enhance portfolio diversification and risk-adjusted returns. Various deep neural networks, including Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Multi-Layer Perceptrons (MLP), and Radial Basis Function Neural Networks (RBFN), are employed to rank stocks based on risk and return characteristics. The top-ranked stocks with the lowest risk are selected for portfolio construction. The entropy constraint is introduced to prevent excessive weight concentration, ensuring a well-diversified portfolio. Historical datasets from the Bombay Stock Exchange (BSE), India, B3 Stock Exchange, Brazil, and Shanghai Stock Exchange, China, are used for validation, with performance assessed on an out-of-sample dataset. Additionally, the efficacy of the suggested approach is evaluated by contrasting it with other machine learning and conventional portfolio optimization techniques. Experimental results demonstrate that the LSTM+MCVaR model with entropy constraint consistently outperforms other deep learning and conventional optimization methods, achieving superior cumulative returns and Sharpe ratios. The findings highlight the potential of combining LSTM forecasting with MCVaR optimization and entropy regularization for robust, diversified portfolio construction.}, archive = {J_EAAI}, author = {Jyotirmayee Behera and Pankaj Kumar}, doi = {10.1016/j.engappai.2025.112198}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112198}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Optimizing mean conditional value-at-risk portfolios through deep neural network stock prediction}, volume = {161}, year = {2025}, }
• (2025). Design and analysis of an unbiased intelligent recommendation system for all-rounders in cricket based on multiple criteria decision making. EAAI, 161, 112197. (www)

  @article{ , abstract = {Designing and analyzing an unbiased intelligent all-rounder recommendation system in cricket is a critical and complex decision-making task, where performance prediction itself is a crucial issue. The majority of existing works on cricket focus on batsmen, bowlers, and teams. However, performance analyses of all-rounders are hardly found. Hence, the motivation of this paper is to propose an artificial intelligence (AI)-based method for assessing the performance of all-rounders utilizing various multiple-criteria decision-making (MCDM) techniques. With this goal in mind, effective attributes are considered for evaluating all-rounder bowling and batting performances. Case studies using a set of 20 all-rounders have been taken from the recent International Cricket Council (ICC) all-rounders list to determine their ranking. The performance of various MCDM techniques is investigated using ICC rankings, with the Spearman Rank Correlation Coefficient applied to the One Day International (ODI) format. The results indicate that the Criteria Importance Through Intercriteria Correlation (CRITIC)-VlseKriterijumska Optimizacija I Kompromisno Resenje (VIKOR) method yields quite promising insights, achieving a higher correlation with ICC rankings (Spearman Rank Correlation: 0.794) and hence can be used as an intelligent AI-based all-rounder recommendation system. Finally, we have also conducted a sensitivity analysis on the ranking outcomes of all-rounders to examine the utility and robustness of our proposed MCDM approach. The proposed approach is beneficial for team selectors, analysts, and fantasy sports platforms, offering a fairer and more reliable ranking of all-rounders.}, archive = {J_EAAI}, author = {Nayan Ranjan Das and Imon Mukherjee and Goutam Paul}, doi = {10.1016/j.engappai.2025.112197}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112197}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Design and analysis of an unbiased intelligent recommendation system for all-rounders in cricket based on multiple criteria decision making}, volume = {161}, year = {2025}, }
• (2025). Semantic analysis-based recommender system using sequential clustering and convolutional neural network. EAAI, 161, 112196. (www)

  @article{ , abstract = {Accurate prediction of user preferences and generation of personalized recommendations remain as critical challenges in intelligent recommendation systems. In this study, we propose a novel recommendation model that transforms the rating prediction problem into a single-label multiclass classification task. The model integrates three key components: (1) ordered clustering information derived from user review text similarity, (2) rating rank similarity reflecting users’ behavioral tendencies, and (3) a convolutional neural network (CNN) to extract semantic representations from user textual data. First, user review embeddings are clustered to capture high-level semantic preferences, where cluster indices are utilized as ordered categorical features. Second, rating rank similarity features are constructed by comparing the relative ranking of items rated by similar users. These features are fused and fed into a CNN model, which outputs a predicted rating class (e.g., 1–5 stars) for each unobserved item, treated as a single-label classification target. To generate final Top-N recommendations, we further incorporate user-specific rating habits and item popularity to re-rank the classification outputs. The experimental results on public benchmark datasets indicate that our model substantially improves the prediction accuracy and recommendation quality compared with existing baselines. The proposed method offers a robust and interpretable approach to bridging textual review semantics, user behavior, and deep learning for rating-aware personalized recommendation.}, archive = {J_EAAI}, author = {Yanjun Xu and Chunqi Tian and Wei Wang and Lizhi Bai}, doi = {10.1016/j.engappai.2025.112196}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112196}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Semantic analysis-based recommender system using sequential clustering and convolutional neural network}, volume = {161}, year = {2025}, }
• (2025). Ranking-oriented cross-modal hashing. EAAI, 161, 112195. (www)

  @article{ , abstract = {Deep cross-modal hashing has become a mainstream solution for multimedia retrieval. Most methods utilize pair-wise or multi-wise loss as proxies for ranking consistency, focusing primarily on local relational patterns. However, these schemes are not well aligned with global ranking metrics, which consider the relative order among all candidate items. This misalignment may push semantically similar items down the list. Additionally, such methods heavily rely on the selection of negative samples, which can cause training instability and require careful tuning. To address this problem, we propose a novel Ranking-Oriented Cross-Modal Hashing (ROCMH) framework, which performs global ranking optimization based on a vision-language model. Specifically, we design a differentiable ranking-oriented objective, called cross-modal ranking alignment loss. It smoothly simulates discrete ranking and naturally models ranking relations in candidate lists, thereby promoting more consistent cross-modal ranking. Meanwhile, considering the perceptual gap between visual and textual, we suggest a nanoparam training strategy with modality-aware prompts. This strategy greatly reduces the number of trainable parameters while ensuring that each modality provides unique signals for semantic information learning. Extensive experiments on three public datasets demonstrate that the method achieves superior retrieval performance and reduces the number of trainable parameters to less than 0.2% of the vision-language model. The source code is available: https://github.com/QinLab-WFU/ROCMH .}, archive = {J_EAAI}, author = {Yadong Huo and Qibing Qin and Wenfeng Zhang and Lei Huang}, doi = {10.1016/j.engappai.2025.112195}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112195}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Ranking-oriented cross-modal hashing}, volume = {161}, year = {2025}, }
• (2025). Matching quality-guided model-free satellite pose estimation. EAAI, 161, 112194. (www)

  @article{ , abstract = {With the development of learning-based techniques and large-scale public datasets, satellite pose estimation has seen significant progress for the past several years. However, most of current methods still rely on a known three-dimensional (3D) model of the object for the pose estimation, limiting its generalization and wide application. To this end, we propose a model-free pose estimation method, which takes as input only a set of images. The proposed method consists of two stages, i.e. the reconstruction stage and pose estimation stage, of which the former reconstructs a 3D model from the input images and the pose is estimated via two-dimensional (2D)-3D feature matching by the latter. More importantly, a matching quality guidance strategy is introduced to further improve the robustness to in-plane rotation during feature matching. Additionally, since no known 3D model is assumed, our method generalizes well to novel objects without retraining. We provide evaluation results on datasets of several satellites with different structures, which demonstrate impressive performances without known models against current methods.}, archive = {J_EAAI}, author = {Zhaoshuai Qi and Yating Liu and Yanning Zhang}, doi = {10.1016/j.engappai.2025.112194}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112194}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Matching quality-guided model-free satellite pose estimation}, volume = {161}, year = {2025}, }
• (2025). Prebuilt spatiotemporal index: An exploration of efficient real-time data storage in intelligent transportation systems. EAAI, 161, 112193. (www)

  @article{ , abstract = {With the widespread deployment of Internet of Things (IoT) devices in Intelligent Transportation Systems (ITS), real-time spatiotemporal data has grown rapidly. Hybrid data-bases have emerged as the mainstream solution for managing such data. The LSM R*-tree, which integrates Log-Structured Merge-trees (LSM-trees) for time-series ingestion and R*-trees for spatial queries, is now a widely used spatiotemporal index. However, storage schemes based on LSM R*-tree structures involve real-time index construction, leading to significant overhead and write latency. To address these challenges, this paper proposes a prebuilt index-based data storage workflow that shifts index construction ahead of data writing. This approach allows the database to directly apply the prebuilt index at write time, thereby minimizing real-time construction costs and enhancing write performance. To support index prebuild, we propose a lightweight embedding scheme, Index2Vec, specifically designed for R*-tree structures. Based on this, we extend the Transformer architecture and develop the R*-tree Prediction Network for ITS (RTPN4ITS), which achieves efficient inference on resource-constrained edge devices. Experimental results show that the prebuilt R*-tree index improves query efficiency by up to 90% over time-index-only schemes and enhances write performance by nearly 50% compared to real-time indexing. The proposed RTPN4ITS model achieves robust accuracy across varying traffic densities, reaching 85% accuracy in dense conditions. Moreover, the Index2Vec embedding enhances the Transformer’s structural awareness. In summary, this paper proposes an efficient prebuilt indexing strategy and lightweight embedding-based model for real-time spatiotemporal data management in ITS.}, archive = {J_EAAI}, author = {Yiran Shao and Kangshuai Zhang and Yong Zhou and Zhenwu Chen and Yang Yang and Lei Peng}, doi = {10.1016/j.engappai.2025.112193}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112193}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Prebuilt spatiotemporal index: An exploration of efficient real-time data storage in intelligent transportation systems}, volume = {161}, year = {2025}, }
• (2025). An error complementarity-based iterative learning approach via categorical boosting for student performance prediction. EAAI, 161, 112192. (www)

  @article{ , abstract = {Predicting student performance is important in achieving academic success and student development in many educational applications (e.g., academic early warning and early interventions). To accurately predict student performance, we propose an error complementarity-based iterative learning approach for student performance prediction. Our goal is to improve the prediction performance by iteratively reducing the prediction error. Specifically, in model construction, we first select the most important features to train the target estimator. Then, we obtain the errors between the target and predicted values. These errors are used to train the error estimator using the remaining features. Similarly, we iteratively train the model to learn from the errors until the desired conditions are met. In model prediction, we predict the testing sample using the target estimator to obtain the predicted value. Next, we predict the error for the next iteration using the corresponding error estimator. This process is repeated, and the final prediction is obtained by adding the predicted value and the error values. The extensive experiments from different educational datasets show that by using our error complementarity-based iterative learning approach, the proposed model outperforms the competing models for prediction accuracy. Furthermore, statistical testing conducted over 20-run experiments confirms the significant advantage of our proposed model. This suggests the iterative learning is effective for predicting student performance.}, archive = {J_EAAI}, author = {Zongwen Fan and Jin Gou and Cheng Wang}, doi = {10.1016/j.engappai.2025.112192}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112192}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An error complementarity-based iterative learning approach via categorical boosting for student performance prediction}, volume = {161}, year = {2025}, }
• (2025). Federated cross-device cluster dynamic time planning warping algorithm in the industrial internet of things. EAAI, 161, 112191. (www)

  @article{ , abstract = {Production lines and heterogeneous devices generate diverse time-series data in the Industrial Internet of Things (IIoT). This diversity necessitates frequent updates of local models when using federated learning (FL). FL has become widespread in the IIoT, these continual model updates can hinder timely information exchange among cross-industry devices. They constrain the system's ability to handle high concurrency in read and write operations. Therefore, we propose a federated cross-device cluster dynamic time planning warping (Fed-cDTW) tailored for the IIoT. We incorporate each device's historical behavior data by constructing a functional characteristic distance matrix. We design a symmetric time warping lower bound function to measure the similarity among multi-source time-series data. This approach enables an adaptive, dynamically organized industrial device cluster. Meanwhile, we employ the federated dynamic (FedD) dual-weight optimization within the device cluster to enhance the models' generalization and robustness. The experimental results evaluated the performance of the proposed method. We compare our method with baselines, Fed-cDTW improves classification accuracy by 0.8 percentage points (pp), 1.7 pp, 9.6 pp, 14.4 pp, and 49.5 pp. Convergence is accelerated by 0.04, 0.05, 0.35, 0.42, and 0.46. Total communication cost is reduced by 40 megabytes (MB), 56 MB, 84 MB, 120 MB, and 168 MB.}, archive = {J_EAAI}, author = {Yifan Zhao and Shibao Sun and Pengcheng Zhao and Yatong Wang and Jianfeng Liu}, doi = {10.1016/j.engappai.2025.112191}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112191}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Federated cross-device cluster dynamic time planning warping algorithm in the industrial internet of things}, volume = {161}, year = {2025}, }
• (2025). Machine learning-based prediction of compressive strength in sustainable self-compacting concrete. EAAI, 161, 112190. (www)

  @article{ , abstract = {The performance and durability of conventional concrete (CC) is significantly influenced by supplementary cementitious materials (SCMs) and recycled coarse aggregate (RCA). Thus, the intrusion of enrich SCMs with RCA in the cementitious matrix delivers utmost properties. This research focuses on the use SCMs and RCA in the self-compacting concrete (SCC) system and their sustainability in the development of concrete resources. Standard experimental methods for forecasting the compressive strength (CS) of SCC have constraints in terms of efficiency, time consumption, and cost. Thus, its prediction is crucial without the need for laborious experimental procedures. This work employs and evaluates a multiplicity of machine learning (ML) models to predict the CS of the cementitious matrix to tackle these issues. Thus, considered a more effective and cost-saving solution in comparison with traditional approaches. Therefore, ML approaches like, gene expression programming (GEP), decision trees (DT), and support vector regression (SVR) were employed. The performance of the model is evaluated by employing the coefficient of determination (R 2 ), statistics, and uncertainty analysis. Individual Conditional Expectation (ICE), and Partial Dependence Plot (PDP) are used to analyze the effect of parameters on strength. The findings suggest that GEP performs best achieving superior R 2 > 0.90 for the training, validation, and test data sets with a lower error. While, the uncertainty analysis shows that all modeled values lie below the threshold value. The ICE and PDP graphs confirms that cement, age, and water-cement ratio have highly relation to outcomes. The RCA replacement ratio is more important than the CA one, and SCMs play an essential role in the development of concrete compressive strength, although not as much as cement. In addition, SP depicts major contribution to SCC. Moreover, graphical user interface (GUI) is also developed to help users/researcher that will facilitate them to estimate the strength of SCC in practical applications.}, archive = {J_EAAI}, author = {Jingguo Gou and Athar Zaman and Furqan Farooq}, doi = {10.1016/j.engappai.2025.112190}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112190}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Machine learning-based prediction of compressive strength in sustainable self-compacting concrete}, volume = {161}, year = {2025}, }
• (2025). A robust fuzzy twin support vector machine with kernel-target alignment for binary classification. EAAI, 161, 112189. (www)

  @article{ , abstract = {Many algorithms similar to the twin version of the support vector machine and their variants have shown better results in the binary classification of nonlinear data points. But in the presence of outlier and noise, these algorithms exhibit low generalization efficiency. To alleviate this challenge, recently proposed, kernel-target alignment based fuzzy least square twin bounded support vector machine (KTA-FLSTBSVM) used fuzzy membership values and is solved using least squares. Inspired by this strategy, for further improvement, we propose a novel approach called decision support kernel-target alignment based fuzzy least square twin bounded support vector machine (DS-KFIFTBSVM). DS-KFIFTBSVM considers the kernelized fuzzy membership values with the regularized twin support vector machine and solves for linear and nonlinear data points using a functional iterative approach. In DS-KFIFTBSVM, the solution is obtained by solving a linearly convergent iterative scheme rather than solving quadratic programming problems. The proposed DS-KFIFTBSVM offers better generalization efficiency, which has been evaluated using both linear and Gaussian kernels, mostly on artificially developed and publicly accessible datasets with diverse dimensionalities. In terms of various performance evaluation metrics, including specificity, precision, false positive rate, rate of misclassification error, F_score, and geometric mean in linear and non-linear cases, DS-KFIFTBSVM outperforms various baseline approaches. It shows the highest accuracy in various datasets, including 98.1884 % for musk dataset (linear kernel) and 100 % for the glass dataset (Gaussian kernel). Further statistical analysis confirms its classification efficiency.}, archive = {J_EAAI}, author = {Deepak Gupta and Barenya Bikash Hazarika and Umesh Gupta and Witold Pedrycz}, doi = {10.1016/j.engappai.2025.112189}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112189}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A robust fuzzy twin support vector machine with kernel-target alignment for binary classification}, volume = {161}, year = {2025}, }
• (2025). Comprehensive assessment of pore-scale segmentation techniques for image-based multiphase flow characterization in porous media. EAAI, 161, 112188. (www)

  @article{ , abstract = {Evaluation of multiphase fluid flow behavior based on pore-scale imaging is significantly influenced by the accuracy of the segmentation techniques. This study sheds light on the influence of various traditional and Artificial Intelligence (AI) algorithms on the major static, dynamic, volumetric, and topological properties of the porous media. The traditional methods include the Gaussian Mixture Model (GMM), Multi-Otsu’s thresholding (MOT), and Random Walk (RW), and the AI-based techniques include Trainable Weka (TW), as well as seven deep convolutional autoencoder architectures. The assessment was carried out by comparing the calculated phase fraction, contact angle, capillary pressure, effective permeability, and topology for multiphase flow images of Bentheimer sandstone. The results revealed that combining Residual Networks (ResNet) and UNet structures (UResNet) marginally performs better based on the statistical metrics. Compared to ground-truth watershed images, applying other approaches except TW yielded acceptable static and volumetric parameters. The main discrepancies were in contact angles, capillary pressure, and phase topology. These discrepancies also influenced calculated effective permeabilities, yielding unrealistic trends. UResNet provided the best accuracy, considering the watershed the ground truth. MOT and RW showed similar performance, while GMM performed inconsistently throughout the analysis and TW produced acceptable results for certain fractional flows. The segmentation approaches affected topological features most, whereas properties such as phase fractions were less affected. The inconsistent behavior and effectiveness of used algorithms indicate that the analysis of multiphase images necessitates a meticulous choice of segmentation techniques, emphasizing AI-based algorithms that ensure consistency across all evaluations.}, archive = {J_EAAI}, author = {Javad Siavashi and Mohammad Sharifi}, doi = {10.1016/j.engappai.2025.112188}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112188}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Comprehensive assessment of pore-scale segmentation techniques for image-based multiphase flow characterization in porous media}, volume = {161}, year = {2025}, }
• (2025). Land use and land cover change detection from multisource satellite imagery – A hybrid convolutional neural network approach. EAAI, 161, 112187. (www)

  @article{ , abstract = {Remote sensing and satellite imagery analysis enable us to observe and understand environmental changes over time. However, such an endeavor can be complex and time-consuming. To overcome this challenge, we propose a framework for leveraging Google Earth Engine (GEE) to analyze Earth's surface changes over time. Our framework starts by assessing land cover changes using satellite imagery, evaluating the effectiveness of vegetation indices in classification, and generating accurate land cover maps for a specific region. We examine the validity of this framework on the area near Rio Do Sul, located in the Santa Catarina district of Brazil, to classify and recognize the alteration in land use and land cover (LULC) from 2019 to 2023. In this effort, we develop a pre-trained Convolutional Neural Network (CNN) for feature extraction from the Landsat 8, Sentinel-1, and Sentinel-2 images and then link those into the random forest (RF) algorithm for spatial morphology and temporal change logic to map the long-term annual time series and detect changes in the region at hand. The novelty of this study arises from developing a CNN-RF hybrid model and implementing a data preprocessing pipeline, which includes progressive techniques for cloud masking and radiometric calibration, to ensure higher accuracy and reliability in land use and land cover classification. This CNN-RF hybrid model successfully traced the increasing and decreasing rate in the built-up area, water bodies, forest, fallow land, plantations, and grassland with an overall accuracy of 96.39 % and 94.15 % for 2019 and 2023, respectively.}, archive = {J_EAAI}, author = {M.S. Babitha and A. Diana Andrushia and N. Anand and M.Z. Naser and Y. Pari}, doi = {10.1016/j.engappai.2025.112187}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112187}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Land use and land cover change detection from multisource satellite imagery – A hybrid convolutional neural network approach}, volume = {161}, year = {2025}, }
• (2025). Safety monitoring digital twin-based centralized model consolidation mechanism using dynamic node selection for multi-worker safety prediction. EAAI, 161, 112186. (www)

  @article{ , abstract = {The construction industry remains one of the most hazardous sectors, requiring innovative solutions to safeguard workers, especially in complex, dynamic outdoor environments. Digital Twin (DT) technology offers promising capabilities for real-time safety monitoring through virtual replicas of physical systems. However, existing DT frameworks rarely integrate comprehensive safety monitoring via a centralized model consolidation mechanism, such as Federated Learning (FL), which is explicitly tailored for multi-worker scenarios. Addressing this challenge, this paper proposes a FL-based Safety Monitoring Digital Twin (SMDT) framework designed to enhance multi-worker safety in resource-constrained settings. This enables real-time safety monitoring and control by representing on-site workers as virtual objects within a synchronized DT environment. A dynamic node selection mechanism based on client performance is employed to optimize global model convergence in FL. To validate the proposed approach, an edge computing-based experimental testbed using actual Raspberry Pi devices was implemented, using real-world construction safety data including worker status, weather conditions, and building structural parameters. Experimental results demonstrate the effectiveness of the proposed framework in significantly improving safety predictions and real-time monitoring efficiency. This research establishes a foundational work towards safer construction sites through intelligent, synchronized safety monitoring systems.}, archive = {J_EAAI}, author = {Sa Jim Soe Moe and Atif Rizwan and Anam Nawaz Khan and Rongxu Xu and Do Hyeun Kim}, doi = {10.1016/j.engappai.2025.112186}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112186}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Safety monitoring digital twin-based centralized model consolidation mechanism using dynamic node selection for multi-worker safety prediction}, volume = {161}, year = {2025}, }
• (2025). Universal entity linking. EAAI, 161, 112185. (www)

  @article{ , abstract = {Entity linking is a process of connecting mentions of entities in a document to corresponding entries in a knowledge base. Traditional models for entity linking often require specific fine-tuning to work with knowledge bases other than those they were originally pretrained on, which limits their flexibility and scalability. Building on the concept of entity profile generation, we propose a novel approach that enables entity linking across various knowledge bases without the need for such fine-tuning. Our pipeline leverages a fine-tuned Large Language Model, a generic embedding model, and a vector store to achieve high precision on the TweekiGold and Reuters-128 datasets. Additionally, it demonstrates strong retrieval rates across the TweekiGold , Reuters-128 , and ISTEX-1000 Wikidata entity linking datasets. We also illustrate the applicability of our method to other knowledge bases, using the Agrovoc knowledge base as an example. This solution offers a more versatile and scalable approach to entity linking.}, archive = {J_EAAI}, author = {Adam Aron Rynkiewicz and Raul Palma and Piotr Formanowicz}, doi = {10.1016/j.engappai.2025.112185}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112185}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Universal entity linking}, volume = {161}, year = {2025}, }
• (2025). A search method for fractured-vuggy reservoir inter-well connectivity path based on multi-modal multi-agent. EAAI, 161, 112184. (www)

  @article{ , abstract = {The complex geological structure of carbonate reservoirs and the intricate fracture-vuggy configurations obscure inter-well connectivity, making its evaluation challenging. Conventional studies primarily rely on seismic static data to delineate fracture-vuggy reservoirs, but the limited recognition accuracy hampers the precise characterization of inter-well connectivity and the spatial configuration of fractures and vugs. To address this, this study constructs a 3D (Three-Dimensional) search environment and use multi-modal static and dynamic data and proposes a multi-agent connected channel search model based on deep reinforcement learning. The model treats multiphase fluid as an agent and incorporates Swin Transformer (Shift Window Transformer) to extract large-scale fracture features from seismic data, providing global prior information for path search. A Graph Attention Network is established based on dynamic response relationships to extract spatial geological features, while a multi-head self-attention mechanism captures real-time fluid interactions in various directions. The model fuses multi-modal features, including seismic attributes and production data, to generate decisions and automatically search for inter-well connectivity channels. Experiments were conducted using the WE1 and WE5 well groups from the fault-controlled karst reservoirs in the Tahe oilfield, with results compared against tracer tests. The findings demonstrate that the proposed model's automatic search paths closely align with seismic data and tracer test results, effectively capturing the spatial distribution of fractures and vugs across different scales. This validates the model's effectiveness in evaluating inter-well connectivity in complex carbonate reservoirs.}, archive = {J_EAAI}, author = {Wenbin Jiang and Dongmei Zhang and Hong Cao and Xiaofeng Wang}, doi = {10.1016/j.engappai.2025.112184}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112184}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A search method for fractured-vuggy reservoir inter-well connectivity path based on multi-modal multi-agent}, volume = {161}, year = {2025}, }
• (2025). Continuous spatio temporal prompts for visual tracking. EAAI, 161, 112183. (www)

  @article{ , abstract = {Currently, visual single-object tracking methods utilize online template updates to combine temporal information. However, these methods rely on confidence scores to evaluate the reliability of the current template, which may result in a template not being updated for an extended period. Moreover, advanced trackers select bounding boxes based solely on the similarity between the template and the search area, which can lead to tracking drift when encountering deformable or similar targets. To alleviate these limitations, we propose a Spatio Temporal Prompt Tracker (STPTrack), which utilizes the prior information about small changes of object state between successive frames. Different from previous tracking methods that mainly rely on templates and similarity scores, STPTrack transfers the object position and shape information of the previous frame to the current frame as continuous spatio temporal prompt for the first time, and realizes the efficient fusion of spatio temporal information through the prompt encoder and the fusion decoder module. Specifically, it encodes the bounding box coordinates or mask information of the previous frame and the response points of the current frame as prompt features, and then combines prompt tokens with search tokens through the fusion decoder to provide the potential location of the object for the search feature map. Our STPTrack sets a new state-of-the-art performance on six tracking benchmark datasets.}, archive = {J_EAAI}, author = {Meng Sun and Xiaotao Liu and Yifan Li and Hongyu Wang and Dian Yuan and Jing Liu}, doi = {10.1016/j.engappai.2025.112183}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112183}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Continuous spatio temporal prompts for visual tracking}, volume = {161}, year = {2025}, }
• (2025). Length-aware center loss for sequence to sequence thai scene text recognition. EAAI, 161, 112182. (www)

  @article{ , abstract = {Thai scene text recognition is a challenging task because Thai can be written in both horizontal and vertical directions, allowing characters to be stacked vertically. To address this issue, our previous work combined vertically stacked characters to create new characters. However, this strategy introduced many similar characters. In this paper, we further investigate this problem and propose the Length-aware Center Loss (LC) for Thai scene text recognition. The original center loss was designed for single object recognition tasks. When applied to multi-label tasks like text recognition, center loss is only effective when the lengths of the labels and prediction results are consistent. This can lead to an extreme case where all images receive incorrect predicted text lengths to minimize loss, severely interfering with the recognition process. Therefore, we propose the Length-aware Center Loss for text recognition. We also design the Length Supervision Module (LSM) and the Feature Clustering Module (FCM) to work alongside the LC loss. LSM predicts text length to provide additional supervision signals, while FCM aims to improve recognition performance by minimizing the distance between the features of corresponding class centers. Since there is no publicly available Thai scene text dataset, we have collected a new dataset containing more than 170,000 samples. Extensive experiments conducted on this dataset show that our method achieves superior performance in both string-level and character-level accuracy compared to other methods.}, archive = {J_EAAI}, author = {Hongjian Zhan and Chun Li and Bing Yin and Yue Lu}, doi = {10.1016/j.engappai.2025.112182}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112182}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Length-aware center loss for sequence to sequence thai scene text recognition}, volume = {161}, year = {2025}, }
• (2025). Diffusion-based vision-language model for zero-shot anomaly detection in medical images. EAAI, 161, 112181. (www)

  @article{ , abstract = {With the rapid advancement of diagnostic technology, the ability to detect pathological areas such as tumors and polyps has significantly improved. This progress provides medical imaging specialists with more precise visual information to support anomaly identification, diagnosis, treatment planning, and patient monitoring. However, existing unsupervised and semi-supervised anomaly detection methods struggle with data privacy constraints, limited annotated medical datasets, and challenges in generalization. Zero-Shot Anomaly Detection (ZSAD), which enables the detection of unseen categories without requiring class-specific training, has emerged as a promising solution by leveraging the vision-language alignment capabilities of Vision-Language Models (VLMs), such as Contrastive Language-Image Pretraining (CLIP). Despite recent progress, ZSAD remains hindered by high noise levels, sparse targets, and poor adaptability in complex medical imaging scenarios. To address these issues, we propose a novel framework: DiffusionCLIP, a diffusion-based VLM for zero-shot anomaly detection in two-dimensional medical images. Specifically, DiffusionCLIP integrates diffusion models into the VLM to progressively denoise multi-level features extracted from the CLIP visual encoder, enhancing feature robustness and discriminability. A multi-level feature fusion strategy is designed to aggregate multi-scale representations from different depths of the visual encoder, ensuring complementary semantic alignment across layers. In addition, a dynamically modulated weight loss function is introduced to adaptively balance the learning of hard and easy samples, further improving model generalization. Extensive experiments on multiple benchmark medical imaging datasets, demonstrate that the proposed method significantly outperforms existing zero-shot anomaly detection approaches in terms of accuracy, robustness, and generalization.}, archive = {J_EAAI}, author = {Yanhui Chen and Hongkang Tao and Zan Yang and Yunkang Cao and Chen Jiang and Longhua Hu and Pengwen Xiong and Haobo Qiu}, doi = {10.1016/j.engappai.2025.112181}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112181}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Diffusion-based vision-language model for zero-shot anomaly detection in medical images}, volume = {161}, year = {2025}, }
• (2025). Surface roughness prediction in turning processes for grey cast iron: A hybrid machine learning approach integrating infrared thermography. EAAI, 161, 112180. (www)

  @article{ , abstract = {Workpiece surface quality is a critical control parameter in machining processes, influencing functional performance, dimensional precision, and wear resistance. However, accurately predicting surface roughness is complex, often limited by the computational demands of traditional high-precision methods and the reliance of existing models solely on cutting parameters, hindering real-time monitoring. This research introduces a hybrid Artificial Neural Network (ANN) methodology specifically developed for predicting surface roughness of grey cast iron GG-25 workpieces machined in turning processes, a material previously unstudied in this context. The methodology integrates real-time infrared thermal measurements from multiple defined regions of interest (ROIs) within the tool-workpiece contact zone, along with cutting parameters. Experimental results demonstrated that feed rate (f) is the most significant cutting parameter (effect = 0.43) affecting surface quality, followed by its combination with cutting speed (V c ) (effect = −0.25) and cutting speed (effect = 0.18). Correlation and non-linear regression analyses revealed complex, often exponential relationships between temperature and surface roughness, showing temperature an upward trend as machining progressed. The developed ANN achieves a correlation coefficient (R) value of 0.99 both when predicting the roughness arithmetic mean deviation (Ra) parameter in training conditions and when using data from experiments not used in training (validations data). Moreover, the model reaches a correlation coefficient value of 0.85 (test data) under cutting conditions different from those used in experiments, demonstrating robustness, significantly outperforming Support Vector Regression (SVR). This model represents a highly potential tool for real-time online inspection.}, archive = {J_EAAI}, author = {Sergio Aguado and Marcos Pueo and Raquel Acero and Ana Cristina Majarena and Jorge Santolaria}, doi = {10.1016/j.engappai.2025.112180}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112180}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Surface roughness prediction in turning processes for grey cast iron: A hybrid machine learning approach integrating infrared thermography}, volume = {161}, year = {2025}, }
• (2025). Legendre multiwavelet-based feature attention guidance lightweight network for accurate steel surface defect classification. EAAI, 161, 112179. (www)

  @article{ , abstract = {Steel surface defects can significantly affect the quality and appearance of industrial products such as aerospace, construction application fields, and so on. Due to the multi-scale morphology, low contrast, and random positions of the defects, achieving a favorable balance between detection accuracy and speed remains challenging in practical applications. To address these challenges, this paper combines Legendre multiwavelet (LW) with feature attention guidance (FAG) mechanism to devise a novel high-accuracy lightweight network (LWFAG-LNet) for surface defect classification. More precisely, LW bases with rich regularities are first utilized to match complex geometric characteristics across multi-wavelet and multi-scale resolution levels. Subsequently, the FAG module effectively fuses shallow high-resolution detailed features with deep low-resolution contextual features, significantly reducing the depth of the convolutional neural network (CNN). To the third step, a lightweight CNN module with four convolutional blocks is designed to further extract deep features while preserving the most valuable defect information. The proposed model achieves the highest recognition accuracy with a simple structure and a compact parameter configuration. Extensive experiments are conducted on the Northeastern University-Classification (NEU-CLS), Xsteel surface defect dataset (X-SDD), and Kungliga Tekniska Högskolan Royal Institute of Technology Textures under varying Illumination, Pose and Scale (KTH-TIPS) dataset to verify the model's effectiveness and generalization capability. The results reach classification accuracies of 99.80 %, 98.49 %, and 99.06 %, outperforming existing models by about 1.45 %, 3.45 %, and 2.04 %, respectively. In summary, the proposed LWFAG module can be flexibly applied to various lightweight network frameworks, demonstrating great potential for real-time surface defect detection.}, archive = {J_EAAI}, author = {Xiaoyang Zheng and Weishuo Liu and Yan Huang}, doi = {10.1016/j.engappai.2025.112179}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112179}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Legendre multiwavelet-based feature attention guidance lightweight network for accurate steel surface defect classification}, volume = {161}, year = {2025}, }
• (2025). Detecting object hue bin in the CIECAM02 uniform color space using a multi-layer perceptron for optimal spectrum generation. EAAI, 161, 112178. (www)

  @article{ , abstract = {In this paper, we present an innovative enhancement to an existing multi-channel light-emitting diode luminaire by integrating artificial intelligence and image analysis capabilities. The system captures an image of the object to be illuminated using a custom-developed Android application. The dominant color in the image is first identified, and then a multi-layer perceptron neural network classifies it into one of 16 hue bins within a uniform color appearance model. This classification process enables the luminaire to select a predefined spectrum designed to enhance the visual appeal of the illuminated object. To ensure robust hue bin classification, we developed a rich dataset with 7920 samples, collected using eleven smartphones under ten distinct light sources. Using this diverse dataset, thirteen multi-layer perceptron neural networks with varying input features were trained and compared, achieving hue classification accuracies, precision, and recall all ranging from 95 % to 99 % on the test set. One of the trained multi-layer perceptron networks was integrated into the Android application, with an inference time of 90 μs on a mid-range smartphone, demonstrating the practical application of artificial intelligence for on-demand hue analysis and lighting optimization. This advancement not only enhances the luminaire's functionality but also enriches the user experience by providing intelligent, on-demand lighting solutions tailored to the object's dominant color.}, archive = {J_EAAI}, author = {Esmat Kishani Farahani and Kaveh Ahmadian Tazehmahaleh}, doi = {10.1016/j.engappai.2025.112178}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112178}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Detecting object hue bin in the CIECAM02 uniform color space using a multi-layer perceptron for optimal spectrum generation}, volume = {161}, year = {2025}, }
• (2025). Generalized zero-shot fault diagnosis method for chillers based on cross-modal information compression. EAAI, 161, 112177. (www)

  @article{ , abstract = {In practical chiller fault diagnosis applications, target fault training data is often inaccessible, which poses significant challenges to the effectiveness of data-driven diagnostic approaches. Generalized zero-shot fault diagnosis (GZSFD) aims to detect and classify all fault types without relying on fault samples from all categories during training. GZSFD models are trained exclusively on data from seen faults—those with available historical data—making them prone to bias toward these seen classes during inference. However, existing methods fail to exploit the fine-grained semantics of fault attributes, resulting in suboptimal performance when addressing feature bias toward seen classes. In this article, an innovative GZSFD framework for chillers based on cross-modal information compression is proposed to overcome this difficulty. This method constructs a cross-modal feature fusion attention (CmFFA) and information compression module, and integrates them into a variational autoencoder with generative adversarial network (VAEGAN), and establishes a CmFFA-VAEGAN network to synthesize high-quality unseen fault samples. Specifically, the CmFFA module effectively aligns local regions of the virtual samples with key words in the textual attributes, enabling the model to attend to fine-grained semantic information and robustly fuse features from different modalities. This facilitates the generation of virtual samples that more accurately reflect the characteristics of the target modality. In addition, an information bottleneck (IB) layer is introduced at the output of the generator to compress redundant information within the fused features, retaining only the key information most relevant to the data augmentation task. This design enhances cross-modal consistency between the synthesized samples and their corresponding attributes, while alleviating the feature shift of unseen class samples toward seen categories. Extensive experiments are designed and executed on the chiller dataset. Experimental results demonstrate that the proposed framework effectively mitigates the bias of synthesized unseen fault samples toward seen categories, leading to a significant improvement in diagnostic accuracy.}, archive = {J_EAAI}, author = {Kexin Jiang and Xuejin Gao and Huayun Han and Huihui Gao and Yongsheng Qi and Xi Zhang and Liang Zhao}, doi = {10.1016/j.engappai.2025.112177}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112177}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Generalized zero-shot fault diagnosis method for chillers based on cross-modal information compression}, volume = {161}, year = {2025}, }
• (2025). Enhancing rotation you only look once version 8 for accurate detection of arbitrarily-oriented and multi-scale construction material in complex environment. EAAI, 161, 112176. (www)

  @article{ , abstract = {Construction material detection is integral to effective material management. While deep learning-based methods have advanced automatic detection, challenges such as arbitrarily oriented objects, similar features, large object scale variations, and noise interference still limit accuracy. This study proposes an enhanced detection method based on improved rotation you only look once version 8, incorporating three key innovations. First, the large selective kernel network employs a spatial selection mechanism to dynamically adjust the network's receptive field, capturing key features and contextual information of various materials in complex construction scenarios. This enhances object detection performance in feature-similar environments. Second, the poly kernel inception network combines non-dilated multi-scale convolutions to extract dense texture features from differently sized objects, while using contextual anchor attention to capture long-range information for small objects. These components work together to improve multi-scale object detection. Third, the rectangular self-calibration module uses horizontal and vertical pooling to model rectangular attention regions, capturing axial global context. Its shape self-calibration function adjusts these regions to better fit arbitrarily oriented construction materials, enhancing focus on objects while suppressing noise interference. Experimental results show a mean average precision of 0.871–2.7 % higher than the baseline model and surpassing other state-of-the-art methods—while maintaining real-time detection speed at 28.3 frames per second. The proposed method improves material detection accuracy in complex construction environments, enabling refined material management on-site. This supports material entry and exit tracking, on-site usage monitoring, inventory management, and procurement planning, while also strengthening lean control of construction progress and costs.}, archive = {J_EAAI}, author = {Yujie Lu and Yuanjun Nong and Dayu Zhu and Shuo Wang}, doi = {10.1016/j.engappai.2025.112176}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112176}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing rotation you only look once version 8 for accurate detection of arbitrarily-oriented and multi-scale construction material in complex environment}, volume = {161}, year = {2025}, }
• (2025). Long-term forecasting of monthly reservoir inflow using deep and machine-learning-based algorithms. EAAI, 161, 112175. (www)

  @article{ , abstract = {This study provides a method for dynamic flow forecasting to detect how to move between periods with high and low-flow peaks. In other words, the model continues by gaining a deep insight into the long-term stable changes of the time series. Four-time horizons (5, 10, 15 and 20 years) were chosen to forecast the streamflow. These time horizons represent the ability of this method to detect the continuation of its movement in the short, medium, and long term. In this study, a dynamic and recurrent cycle is used to forecast river streamflow based on long-short-term memory (LSTM) and artificial neural network (ANN) models. The 12-month time delay of precipitation and streamflow of the Zayanderud river from 1970 to 2019 was used as input variables for training and validation of the models. In each stage of the forecasting process, 1 year (12 months) is forecasted and used as input for the next stage. The results show that the ANN model performs better in shorter time horizons and the LSTM model in longer time horizons, respectively. The best performance of the ANN model occurs in the 10-year forecast (Mean Absolute Percentage Error (MAPE) = 44.3, Nash–Sutcliffe efficiency coefficient (NSCE) = 0.58, Root Mean Square Error (RMSE) = 43.75) and the best performance of the LSTM model occurs in the 15-year forecast (MAPE = 48.9, NSCE = 0.54, RMSE = 47.12). Finally, the selected period is used to forecast the streamflow in future periods when observational data are not available.}, archive = {J_EAAI}, author = {Bahram Ghenaati and Mohammad Reza Nikoo and Mohammad G. Zamani}, doi = {10.1016/j.engappai.2025.112175}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112175}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Long-term forecasting of monthly reservoir inflow using deep and machine-learning-based algorithms}, volume = {161}, year = {2025}, }
• (2025). Characteristics detection of surface agglomerated abrasive on electroplated diamond wire: A measurement method based on semantic segmentation. EAAI, 161, 112173. (www)

  @article{ , abstract = {Electroplated Diamond Wire (EDW) serves as the primary tool for cutting semiconductor materials. Abrasive agglomeration on its surface induces sudden fluctuations in sawing force, exacerbates material surface damage, and shortens tool lifespan. However, existing testing methods primarily focus on abrasive counts while generally lacking effective means to quantify key characteristics of agglomerated abrasives (e.g., density, maximum size, and number of abrasives contained), leading to substantial limitations in the EDW quality assessment system. To address this critical gap, this study proposes a U-shape Network (UNet)-based semantic segmentation model for EDW surface abrasive grain segmentation, termed Electroplated Diamond Wire U-shape Network (EDW-UNet). By integrating Atrous Spatial Pyramid Pooling (ASPP) and Mamba-like Linear Attention (MLLA), EDW-UNet enhances the detection of multi-scale targets and the accuracy of complex boundary segmentation, enabling precise segmentation of both single and agglomerated abrasives on EDW surfaces. Experimental results demonstrate that EDW-UNet achieves a mean Intersection over Union of 0.917 and a mean Pixel Accuracy of 0.961 on the test set, outperforming the baseline UNet and other mainstream models, with an inference speed of up to 12.82 frames per second. For the first time, a feature extraction algorithm developed from the segmentation results enables measurement of key parameters, including the density, maximum size, and number of abrasives contained in agglomerates. This study not only addresses the technical gap in quantitative detection of agglomerated abrasive features but also provides a novel high-precision, quantitative standard for optimizing EDW manufacturing processes and assessing product quality, thereby enhancing the cutting efficiency and yield of semiconductor materials. The code is available at: https://github.com/huangwenbin123/EDW-UNet .}, archive = {J_EAAI}, author = {Wenbin Huang and Yufei Gao and Shengtan Hu and Dameng Cheng}, doi = {10.1016/j.engappai.2025.112173}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112173}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Characteristics detection of surface agglomerated abrasive on electroplated diamond wire: A measurement method based on semantic segmentation}, volume = {161}, year = {2025}, }
• (2025). Risk-response coupling in underground structures under liquefiable soil conditions: A causality-informed dynamic bayesian network integrated framework. EAAI, 161, 112171. (www)

  @article{ , abstract = {Underground structures in liquefiable soils face complex seismic risks that can trigger cascading failures. This study proposes a Granger causality-informed Dynamic Bayesian network (G-DBN) framework to capture the temporal propagation of seismic risk in such systems. Firstly, a system risk assessment model integrates multiple performance indicators through Cloud Model (CM) to quantify overall risk levels, considering uncertainties associated with soil liquefaction and structural responses. Subsequently, a structural dynamic risk inference model is established using Dynamic Bayesian network (DBN), combining Granger Causality (GC) analysis with engineering-informed relationships to define the network structure. The input features include key structural state variables such as tunnel cross-section convergence ( T 4 ), tunnel uplift displacement ( T 6 ), station uplift displacement ( S 5 ), and inter-story drift angle ( S 6 ), and the aggregated structural risk indicator serves as the target variable. This framework enables the temporal propagation of risk across interconnected structural nodes, and elucidates the mechanisms by which liquefiable soil deformations and structural responses interact within the soil-structure system. Results showed that the risk characteristic value (Expectation , E x ) decreased from 29.12 % (percentage) to 5.21 % as the Peak Ground Acceleration (PGA, expressed in units of gravitational acceleration g) increased from 0.1 g to 0.7 g. The proposed G-DBN model demonstrates robust predictive capabilities, achieving coefficient of determination ( R 2 ) values exceeding 0.95 across multiple seismic intensity conditions. Additionally, tunnel cross-section convergence ( T 4 ) was identified as the most critical factor affecting risk propagation in the coupled underground systems. By integrating holistic risk quantification with dynamic propagation analysis, this study offers a robust tool for understanding dynamic risk evolution and supports decision-making for seismic resilience of underground infrastructure in liquefiable soils.}, archive = {J_EAAI}, author = {Heqi Kong and Xiaohua Bao and Jun Shen and Xiangcou Zheng and Xiangsheng Chen}, doi = {10.1016/j.engappai.2025.112171}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112171}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Risk-response coupling in underground structures under liquefiable soil conditions: A causality-informed dynamic bayesian network integrated framework}, volume = {161}, year = {2025}, }
• (2025). Neural operator-enabled forward and inverse modeling of laser-induced surface acoustic waves and applications in nondestructive evaluation. EAAI, 161, 112170. (www)

  @article{ , abstract = {Laser-induced surface acoustic wave (SAW)-driven nondestructive evaluation offers high-resolution, non-contact characterization of subsurface microstructures. However, its practical application is often limited by the high computational costs associated with traditional numerical simulation methods. Recently, machine learning has emerged as an attractive alternative to accelerate these simulations. In this paper, we develop a neural operator-enabled framework for both forward and inverse modeling of laser-induced SAW propagation. A general dataset with randomly generated subsurface structures is used to evaluate and quantify the model's performance in both wave propagation and subsurface inversion problems. Three potential applications are then investigated: subsurface crack localization, multilayer structure characterization and polycrystalline grain imaging. The results demonstrate that the neural operator-enabled model achieves satisfactory accuracy even in the presence of noise and source waveform variations, underscoring its potential as an efficient and accurate surrogate model for practical nondestructive evaluation using laser-induced SAWs.}, archive = {J_EAAI}, author = {Zaiwei Liu and Zhongqing Su}, doi = {10.1016/j.engappai.2025.112170}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112170}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Neural operator-enabled forward and inverse modeling of laser-induced surface acoustic waves and applications in nondestructive evaluation}, volume = {161}, year = {2025}, }
• (2025). A multi-objective double Q-learning-based hyper-heuristic algorithm for aluminum production and transportation integrated scheduling problem. EAAI, 161, 112169. (www)

  @article{ , abstract = {The cooperative process of electrolytic aluminum production and aluminum liquid distribution is a typical distributed integrated scheduling problem. The aluminum production and transportation integrated scheduling problem (APTISP) is a strong coupling problem that is studied with the optimal objectives of total completion time and total energy consumption. A mathematical model is constructed for the APTISP. A multi-objective double Q-learning-based hyper-heuristic (MDQHH) algorithm is designed to solve the APTISP. In the MDQHH algorithm, a heuristic rule-based initialization operation is designed to generate the initial population of the APTISP. Nine problem-specific low-level perturbation heuristics are constructed to explore the APTISP solution space. A dynamically adjustable ε -greedy strategy is designed to select the low-level perturbation heuristics, and ε gradually decreases as the learning process progresses. Two Q-tables are utilized alternately to select actions and update the Q-values based on the current solution in each iteration. The experimental results show that the MDQHH algorithm outperforms several state-of-the-art comparison algorithms in addressing the performance of APTISP.}, archive = {J_EAAI}, author = {Fuqing Zhao and Ting Yang and Tianpeng Xu and Jianlin Zhang}, doi = {10.1016/j.engappai.2025.112169}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112169}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-objective double Q-learning-based hyper-heuristic algorithm for aluminum production and transportation integrated scheduling problem}, volume = {161}, year = {2025}, }
• (2025). Scalable multi-agent reinforcement learning for factory-wide dynamic scheduling in semiconductor manufacturing. EAAI, 161, 112168. (www)

  @article{ , abstract = {Real-time dynamic scheduling in modern manufacturing is highly complex due to frequent disturbances and intricate operational constraints. While reinforcement learning (RL) has shown promise, existing approaches often rely on extensive dispatching rules and struggle to scale to factory-wide settings. This study introduces a scalable multi-agent RL (MARL) framework with a leader–follower structure, enabling decentralized agents to handle sub-problems while maintaining global coordination through abstract goals. To further enhance robustness, a limited rule-based conversion algorithm is proposed to mitigate performance degradation from poor agent decisions. Our experimental results demonstrate that the proposed model outperforms the state-of-the-art deep RL-based scheduling methods in various aspects. Additionally, the proposed model provides the most robust scheduling performance to demand changes. Overall, the proposed MARL-based scheduling model presents a promising solution to the real-time scheduling problem, with potential applications in various manufacturing industries.}, archive = {J_EAAI}, author = {Jaeyeon Jang and Diego Klabjan and Han Liu and Nital S. Patel and Xiuqi Li and Balakrishnan Ananthanarayanan and Husam Dauod and Tzung-Han Juang}, doi = {10.1016/j.engappai.2025.112168}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112168}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Scalable multi-agent reinforcement learning for factory-wide dynamic scheduling in semiconductor manufacturing}, volume = {161}, year = {2025}, }
• (2025). Research on global path planning algorithm based on indoor map partition preprocessing. EAAI, 161, 112167. (www)

  @article{ , abstract = {Mobile robots utilize Simultaneous Localization and Mapping (SLAM) technology to generate environmental maps and determine their locations within these environments. Subsequently, they employ path planning algorithms to complete navigation tasks. Although existing path planning algorithms are relatively mature, they still exhibit inefficiencies in complex indoor environments. To address this issue, this paper introduces an Indoor Map Partitioning Preprocessing (IMPP) algorithm, which identifies and segments irregularly shaped, complex rooms to accelerate the path planning process. The method initially utilizes the Robot Operating System (ROS) to construct an indoor map dataset and subsequently applies an image segmentation model to identify and enclose rooms. By combining image processing techniques with path planning algorithms, this method can obtain room index information and successfully exclude irrelevant areas from the path planning process. Ultimately, the IMPP algorithm is integrated with a variety of global path planning algorithms. Experimental results demonstrate that in complex indoor environments, this method significantly surpasses existing partitioning methods in terms of room recognition accuracy. Moreover, it decreases the number of expansion points in global path planning algorithms, significantly enhancing processing speed and efficiency.}, archive = {J_EAAI}, author = {Jifan Yang and Xiaoling Li and Xiaoyang Liu and Xunding Pan and Lei Wang}, doi = {10.1016/j.engappai.2025.112167}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112167}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Research on global path planning algorithm based on indoor map partition preprocessing}, volume = {161}, year = {2025}, }
• (2025). An ensemble transfer learning strategy for short-term electricity load forecasting of data-sparse buildings. EAAI, 161, 112166. (www)

  @article{ , abstract = {The “data silos” caused by the lack of available historical or shared information is an important reason affecting forecasting performance of building’s short-term power demand, thus limits the flexibility of demand response of building group. Knowledge’s transfer learning (TL) from similar domains has been proved an efficient tool to enhance target prediction’s performance. To improve data-sparse building’s energy forecasting accuracy and robustness simultaneously, an ensemble strategy based on different TL models is proposed in this study. Firstly, a two-stage similar data selection method is proposed for source domain construction. The multi-kernel maximum mean discrepancy (MK-MMD) is used for selecting similar buildings to the target one; the nearest neighbor search (NNS) method is adopted for picking the most helpful data from the similar buildings to form the dataset of source domain. To determine the optimal number of source domains for transfer learning, an objective function is also formulated. Then, two effective TL models, i.e., instance-based TL (IBTL) and model-based TL (MBTL) are both built as sub models. On this basis, a parallel ensemble framework is designed with model weights adjusted by multiple linear regression. For case study, a public data set containing 1,449 buildings is treated as candidate source domain set for knowledge transfer, and two educational buildings with sparse historical data are treated as target buildings for hourly electricity load forecasting. Forecasting results show that compared with the previous reported TL models, the proposed ensemble strategy always achieves the best performance in two cases and the prediction accuracies are improved by 41.8% and 29.3% (MAPE) respectively. The impact of source domain number and target data amount on prediction performance are also analyzed and discussed in details.}, archive = {J_EAAI}, author = {Xingqiao Liu and Borui Wei and Wenping Xue and Xiaoying Li and Kangji Li}, doi = {10.1016/j.engappai.2025.112166}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112166}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An ensemble transfer learning strategy for short-term electricity load forecasting of data-sparse buildings}, volume = {161}, year = {2025}, }
• (2025). Intelligent multispectral image recognition by using credibility-based square root cubature kalman filters and broad particle swarm learning. EAAI, 161, 112165. (www)

  @article{ , abstract = {This paper proposes intelligent multispectral image recognition by using credibility-based Square Root Cubature Kalman Filters (SRCKF) and broad particle swarm learning to address the issue of low recognition accuracy in multispectral target detection using Kalman filtering. This method initially enhances the estimation accuracy of SRCKF by integrating credibility theory. Secondly, the regression parameters of a Broad Learning System (BLS) are tuned with Particle Swarm Optimization (PSO), forming a PSO-BLS network that offers fast, lightweight feature learning. The PSO-BLS acts as an regressor that compensates residual errors produced by the credibility-based SRCKF, thereby refining the overall filtering precision. Subsequently, a particle swarm-optimized broad learning neural network is employed to rectify the SRCKF results from a compensatory standpoint, thereby enhancing the system’s filtering precision. Ultimately, the adaptive search window generated by Camshift(Continuously adaptive Mean shift) can be employed as an input for the credibility SRCKF, thereby facilitating the acquisition of more reliable real-time observation data. Finally, in the experimental part, the proposed algorithm is validated by the MNIST(Modified National Institute of Standards and Technology) dataset , CIFAR-10(Canadian Institute For Advanced Research) dataset, GTOT(Grayscale-Thermal Object Tracking) dataset and OTB2015(Object Tracking Benckmark 2015) dataset respectively for its generalization ability and robustness in static classification tasks and the effectiveness of dynamic target detection and tracking in real lighting scenes, Occlusion and motion blur scenes, while the performance is more advantageous than some mainstream algorithms.}, archive = {J_EAAI}, author = {Quanbo Ge and Zijian Xue and Hao Wang and Yuanliang Wang}, doi = {10.1016/j.engappai.2025.112165}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112165}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Intelligent multispectral image recognition by using credibility-based square root cubature kalman filters and broad particle swarm learning}, volume = {161}, year = {2025}, }
• (2025). Accidents analysis of mining industry through semantic text representation and dimensionality reduction: An integrated clustering framework. EAAI, 161, 112164. (www)

  @article{ , abstract = {Despite significant improvements in worker health and safety in recent decades, the mining industry still experiences fatal and non-fatal accidents. This underscores the critical need for a more nuanced understanding of accident causation patterns through accident data analysis. While conventional analytical approaches have yielded valuable insights, the extensive information embedded within text-based accident narratives remains underutilized. To address this gap, this study presents an artificial intelligence (AI)-based framework that integrates transformer-based natural language processing, nonlinear dimensionality reduction, and unsupervised machine learning to analyze and cluster accident narratives from the U.S. mining industry. Specifically, this study uses Sentence-BERT (SBERT), a sentence embedding model based on Bidirectional Encoder Representations from Transformers (BERT), to extract the high-dimensional semantical representation of accident narratives. These embeddings are then mapped to a low-dimensional space using the Uniform Manifold Approximation and Projection (UMAP) technique, followed by clustering with the k-means algorithm and subsequent hazard-focused cluster analysis. The primary contribution to AI lies in demonstrating the effectiveness of combining modern sentence embedding techniques with dimensionality reduction and clustering for the semantic analysis of safety-related narratives. From an engineering standpoint, this framework enables the identification of latent accident patterns that can inform hazard detection and guide safety interventions in the mining industry. The resulting clusters reveal diverse accident patterns across mining operations. In clusters associated with underground mining, a high proportion of incidents (ranging from 84 to 98 %) involved no equipment, with distinct injury patterns: torso injuries (67 %) from over-exertion, lower extremity injuries (58 %) from slips/falls, and upper extremity injuries from over-exertion (95 %) and material handling (85 %). Equipment-related clusters revealed strong associations with drilling tools (92 %), loaders (98 %), and bolting equipment (96 %). Clusters associated with strip/quarry/open pit operations exhibited a high frequency of vehicle-related accidents (98 % transportation, 99 % loaders), often resulting in multiple body part injuries. Milling operation clusters show 52–97 % no-equipment accidents, with injury patterns similar to those in underground mining. Additionally, noise-induced hearing loss (96–97 %) was observed in clusters spanning all mining operation types. These findings offer actionable insights for safety professionals and support data-driven, targeted interventions in the mining industry.}, archive = {J_EAAI}, author = {Abid Ali Khan Danish and Snehamoy Chatterjee and Kumar Vaibhav Raj and Rennie Kaunda and Hugh Miller and Aref Majdara}, doi = {10.1016/j.engappai.2025.112164}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112164}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Accidents analysis of mining industry through semantic text representation and dimensionality reduction: An integrated clustering framework}, volume = {161}, year = {2025}, }
• (2025). A lightweight and improved you only look once model using GhostWise convolution and attention mechanism for accurate plant disease detection. EAAI, 161, 112163. (www)

  @article{ , abstract = {Artificial intelligence enables early plant disease detection to support food security, yet achieving high accuracy with low computational cost on multi-class datasets remains challenging. To bridge this gap and solve the problem, we propose an improved version of the “You Only Look Once version 8 nano” (YOLOv8n) model that integrates a convolutional block attention module and custom coarse-to-fine modules into the baseline YOLOv8n architecture. To improve both efficiency and detection performance, the proposed model incorporates a combination of Ghost module and Depthwise convolution for lightweight feature extraction. Additionally, the attention modules emphasize important spatial and channel-wise features, helping the model focus better on critical information. The performance of the proposed model is compared with YOLO models including YOLOv5n, YOLOv5s, YOLOv7, YOLOv8n, and YOLOv8s. The experimental results show that the improved model outperforms the benchmark YOLO models. In addition, the improved model achieves mean Average Precision at Intersection over Union threshold of 0.5 ( mAP@0.5 ) 0.742, representing an improvement of 3.37 % over the baseline YOLOv8n. For the metric mean Average Precision at thresholds from 0.5 to 0.95 ( mAP@0.5:0.95 ), the improved model shows an improvement of 5.13 % over YOLOv8n. The precision measure reveals an improvement of 8.99 % and F1 score of 0.6856 outperforms YOLOv8n by 5.85 %. In addition, generalization experiments are conducted using commonly preferred plant disease datasets. These results validate the robustness, accuracy and stability of the improved model in effectively detecting plant leaf diseases.}, archive = {J_EAAI}, author = {Bunyamin Dikici and Mehmet Fatih Bekciogullari and Hakan Acikgoz and Serkan Ozbay}, doi = {10.1016/j.engappai.2025.112163}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112163}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight and improved you only look once model using GhostWise convolution and attention mechanism for accurate plant disease detection}, volume = {161}, year = {2025}, }
• (2025). Hybrid optimization approach for syngas-fueled gas turbines: Integrating inverse problem solving and machine learning techniques. EAAI, 161, 112162. (www)

  @article{ , abstract = {This study investigates the efficiency of gas turbines powered by syngas derived from gasification, combining a detailed thermodynamic model with a hybrid optimization approach that integrates inverse problem solving via the Levenberg-Marquardt (LM) algorithm and supervised machine learning using Gradient Boosted Trees (GBT). The LM method is used to estimate turbine and compressor performance parameters based on efficiency maps extracted through a custom image processing routine. These maps are applied to calibrate isentropic efficiencies and define the system's operational boundaries. The resulting performance curves enhance the accuracy of the thermodynamic cycle model. To identify optimal operating conditions that maximize thermal efficiency and minimize entropy generation, the inverse problem formulation is integrated with the GBT model, enabling data-driven optimization based on simulated system behavior. All modeling and simulations are conducted in Wolfram Mathematica version 14.0, while verification is performed with the commercial software Gasturb version 14.0, using syngas composition data available in the literature. As a case study to demonstrate the method's applicability, syngas obtained from sewage sludge gasification is analyzed. The findings after optimization indicated an average thermal efficiency of 40 % when using syngas. The analysis revealed that the fuel mass flow rate contributes approximately 45 % to the efficiency gains and more than 70 % to the exergy reduction, while the excess air ratio contributes around 50 % to the cycle's efficiency. The study demonstrates the value of integrating LM-based modeling and GBT optimization, indicating significant potential for enhancing the performance of syngas-fueled turbines. Moreover, it highlights syngas from sewage sludge as a promising sustainable energy source.}, archive = {J_EAAI}, author = {Hiago David Zogbi Silva Oliveira and Vítor Caldeira de Andrada Bastos and York Castillo Santiago and Isabela Florindo Pinheiro}, doi = {10.1016/j.engappai.2025.112162}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112162}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hybrid optimization approach for syngas-fueled gas turbines: Integrating inverse problem solving and machine learning techniques}, volume = {161}, year = {2025}, }
• (2025). Frequency-wavelet adaptive basis network for long-term time series forecasting. EAAI, 161, 112161. (www)

  @article{ , abstract = {Transformer-based models have demonstrated remarkable efficacy in long-term time series forecasting. Nevertheless, their fixed-scale feature extraction and static spectral processing limit adaptability to inherent multi-scale temporal dependencies. Additionally, conventional self-attention mechanisms struggle to effectively model hierarchical temporal structures. To address these limitations, we propose the Frequency-Wavelet Adaptive Basis Network (FWBNet), a novel framework that integrates three synergistic components including a Multi-scale Adaptive Basis Module (MS-ABM) for multi-resolution feature extraction, a Dual-Stream Wavelet Cross-Attention (DWCA) system with learnable wavelet transformations, and a computationally efficient Frequency-domain Multi-Layer Perceptron bottleneck (FreMLP bottleneck) for selective spectral processing. The MS-ABM innovatively integrates multi-scale features of time series through an adaptive basis function selection mechanism, dynamically fusing characteristics at different scales to form hierarchical representations. These multi-dimensional representations are then processed by the DWCA system, which establishes an efficient modeling framework for cross-scale temporal dependencies by combining bi-directional attention with learnable wavelet transforms. Building upon this foundation, the FreMLP bottleneck layer strategically performs selective feature optimization in the frequency domain, extracting key frequency information through spectral sparsification techniques to form an end-to-end multi-scale temporal feature learning pipeline. Experimental evaluations on six benchmark datasets demonstrate that FWBNet significantly outperforms state-of-the-art models, achieving mean squared error (MSE) reductions of 14% for multivariate and 21% for univariate predictions. This study presents a comprehensive framework for effective multi-scale temporal information integration in complex time series forecasting applications.}, archive = {J_EAAI}, author = {Qiang Lai and Yang You}, doi = {10.1016/j.engappai.2025.112161}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112161}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Frequency-wavelet adaptive basis network for long-term time series forecasting}, volume = {161}, year = {2025}, }
• (2025). Balanced sample repository for knowledge distillation in data-free image classification scenario. EAAI, 161, 112160. (www)

  @article{ , abstract = {To address strict data security and privacy requirements in the field of knowledge distillation for image classification, generative network-based data-free knowledge distillation produces fake images for training student models to perform image classification tasks. However, we discover that the datasets consisting of those fake images often suffer from imbalances, such as unclear class margins, uneven intra-class distribution, and skewed label distribution, impeding classification accuracy. To this end, this paper proposes the balanced sample repository method for data-free knowledge distillation in the context of image classification, comprising three components: dynamic pseudo-supervised loss weight (DLW), biased label generator (BLG), and intra-class homogeneity operator (IHO). Specifically, DLW adjusts the pseudo-supervised loss weight to maintain distinct class margins, BLG uses the roulette algorithm to balance label distribution, and IHO employs local outlier factors to remove outliers and homogenize intra-class distributions. The fake images generated by the generative network are initially stored in the sample repository, and those three components achieve a balanced dataset by influencing this sample repository. The experiments conducted across five recognized image classification datasets demonstrate that the proposed method achieves competitive results and produces more balanced samples.}, archive = {J_EAAI}, author = {Yafeng Sun and Xingwang Wang and Junhong Huang and Shilin Chen}, doi = {10.1016/j.engappai.2025.112160}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112160}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Balanced sample repository for knowledge distillation in data-free image classification scenario}, volume = {161}, year = {2025}, }
• (2025). A lightweight detection model for green navel oranges in natural environments. EAAI, 161, 112157. (www)

  @article{ , abstract = {In natural environments, the detection of green navel oranges in the unstructured plantings of Gannan navel oranges presents challenges, including low model accuracy and excessive parameter size. This paper integrates existing technologies to propose a lightweight green tangerine detection model based on You Only Look Once version 5s (YOLOv5s). The aim is to improve the accuracy of green tangerine identification and detection in natural environments through model improvements. Firstly, the original YOLOv5 network model is optimized by replacing the original CSPDarknet53 (Cross Stage Partial Darknet53) backbone network with MobileNetV3 in order to reduce model parameters. Concurrently, the Squeeze-and-Excitation (SE) module in the network is substituted with the Convolutional Block Attention Module (CBAM), resulting in a lightweight network. Secondly, CBAM is integrated into the C3 network (cross-stage partial bottleneck with three convolutional layers) in the neck region to enhance the model's ability to extract green navel orange features after lightweighting. Finally, improve the loss function of the model to reduce the positioning error of the model for low-quality targets. The experimental results demonstrate that the enhanced model attains a 3.6 % enhancement in precision, 2.3 % in recall, and 2.1 % in mean Average Precision (mAP) in comparison with the original YOLOv5 model. The model's size is 3.9 megabytes (MB), representing a reduction of 72.9 % from the original 14.4 MB. Research shows that our method provides an efficient and accurate solution for identifying green navel orange targets in complex environments, with potential application value in agricultural automation for orchard management and monitoring.}, archive = {J_EAAI}, author = {Dongyuan Zhang and Qiusheng Li and Xindi Yuan}, doi = {10.1016/j.engappai.2025.112157}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112157}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A lightweight detection model for green navel oranges in natural environments}, volume = {161}, year = {2025}, }
• (2025). Efficient real-time instance segmentation of garment for intelligent robot tie-dye based on you only look once version 11 network. EAAI, 161, 112156. (www)

  @article{ , abstract = {Tie-dye, as an intangible cultural heritage with a long history, has gained global popularity due to its unique artistic value and versatile craftsmanship. However, its reliance on manual operation leads to low production efficiency and inconsistent quality. To address this, we propose an improved You Only Look Once version-11 (YOLOv11) instance segmentation model, Garment-YOLO, to efficiently segment the garment regions of robotic tie-dye by artificial intelligence technology. First, the C3k2_DynamicMixFormer (C3k2_DMF) module introduces dynamic kernel weight selection mechanism and multi-scale fusion to effectively balance local details and global information. Meanwhile, the Dual-Cross Recalibration Feature Pyramid Network (DCR-FPN) is proposed to enhance the detail preservation of edge region by selectively aggregating semantic and boundary information. Furthermore, the Superior-Head replaces depthwise convolution (DWConv) with part convolution (PConv), significantly reducing model complexity while maintaining performance. Experimental results demonstrate that Garment-YOLO achieves 167.4 frames per second (FPS) and maintains an optimal trade-off between inference speed and segmentation accuracy, reaching 91.9 % mean average precision (mAP)50–95 (Box), 91.5 % mAP50-95 (Mask). To further validate its practical performance, we conducted a 72-h production-line comparison, showing that compared to the baseline, we increase the product qualification rate by 24.6 %, reduce dye waste by 19.4 %, and increase the total production by 52.5 % compared to manual tie-dye. This study not only provides a feasible solution for the intelligent transformation of traditional tie-dye but also contributes to the preservation and development of this intangible cultural heritage. The code will be released on GitHub ( https://github.com/fudifu123/GARMENT-YOLO ). A video that intuitively introduces our research ( https://www.bilibili.com/video/BV13MgczAEkB ).}, archive = {J_EAAI}, author = {Difei Feng and Qihong Zhou and Lei Xiao and Kunfeng Ge and Hangzhou Ma and Li Zhou}, doi = {10.1016/j.engappai.2025.112156}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112156}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Efficient real-time instance segmentation of garment for intelligent robot tie-dye based on you only look once version 11 network}, volume = {161}, year = {2025}, }
• (2025). A deep patch network with spatiotemporal meta-parameter learning for soft sensor modeling of industrial processes. EAAI, 161, 112155. (www)

  @article{ , abstract = {Soft sensing techniques have been widely employed in industrial processes to predict hardly measurable quality variables using easily measurable process variables. Effective modeling strategies for representing spatiotemporal features are highly desired in soft sensing applications for dynamic industrial processes. Deep networks, such as graph convolutional network and gated recurrent unit, are frequently employed to capture the spatial or temporal characteristics of sequential data. Nevertheless, effectively modeling and utilizing spatiotemporal heterogeneity remains a challenging problem. Thereby, a p atch network with spatiotemporal m eta- p arameter l earning (PMPL) is proposed in this paper, which presents an uniform framework of modeling spatiotemporal heterogeneous feature representations for soft sensing. In PMPL, a patching strategy is utilized to segment times series into sub-series-level patches, allowing the retention of local semantic information while capturing long-term dependencies. Next, a novel spatiotemporal meta-parameter learning approach is proposed to generate specific parameters for a spatial encoder, temporal encoder and spatiotemporal decoder through establishing corresponding embedding layers. In this way, the spatial, temporal and spatiotemporal dependencies within the process data can be adaptively captured and comprehensively refined from a uniform perspective. Extensive experiments are conducted based on two real production processes, which illustrate the feasibility and efficacy of the proposed model.}, archive = {J_EAAI}, author = {Xudong Shi and Kangping Du and Weili Xiong and Humberto Morales and Adriana Amicarelli}, doi = {10.1016/j.engappai.2025.112155}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112155}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A deep patch network with spatiotemporal meta-parameter learning for soft sensor modeling of industrial processes}, volume = {161}, year = {2025}, }
• (2025). Deep bioinspired evolutionary stacking algorithm for unpaired multimodal cell classification calibration. EAAI, 161, 112153. (www)

  @article{ , abstract = {Single-modality deep learning approaches for cell image classification exhibit inherent limitations in informational diversity when processing cross-institutional datasets acquired under varied imaging protocols. In contrast, multimodal cell imaging has emerged as a promising alternative for addressing data heterogeneity through comprehensive information integration. This study introduces a novel multimodal alternate training decision-making architecture based on a stacking algorithm for unpaired multimodal cell classification calibration. The method leverages deep bioinspired evolutionary networks combined with kernel-based support vector machines. Specifically, deep base classifiers incorporating multimodal concepts are derived from heterogeneous cell datasets. Each base classifier employs a bioinspired strategy to perform alternate training between two evolutionary densely connected attention networks. To mitigate class imbalance, where diseased cells are significantly outnumbered by normal cells, we incorporate a Shannon entropy loss term. Finally, multiple kernel-based support vector machines serve as meta classifiers, transforming high-level multimodal concepts into a separable feature space for robust decision-making. Experimental results demonstrate the superiority of our approach over existing algorithms for unpaired multimodal cell image classification. Our findings emphasize the importance of alternate training intra-modality classifiers and inter-modality fusion calibration for accurate and reliable medical image classification. All source code for this work will be publicly available on GitHub.}, archive = {J_EAAI}, author = {Lili Zhao and Di Xu and Xueping Tan and Jinzhao Yang and Weiping Ding and Hengde Zhu and Lichi Zhang and Qian Wang}, doi = {10.1016/j.engappai.2025.112153}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112153}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep bioinspired evolutionary stacking algorithm for unpaired multimodal cell classification calibration}, volume = {161}, year = {2025}, }
• (2025). Temporal-frequency joint hierarchical transformer with dynamic windows for speech emotion recognition. EAAI, 161, 112152. (www)

  @article{ , abstract = {Speech Emotion Recognition (SER) aims to identify the emotional state of a speaker from speech signals, serving as a critical prerequisite for achieving natural human–computer interaction. In speech signals, emotional information is inherently distributed across diverse frequency bands and temporal scales, with emotional cues in distinct regions exhibiting varying levels of heterogeneity or interdependence. Existing Transformer-based methods face limitations in precisely localizing salient temporal-frequency regions and modeling their inter-regional relationships. To address these challenges, a temporal-frequency joint hierarchical Transformer with dynamic window mechanisms, abbreviated as TF-DWFormer, is proposed to capture critical emotional cues and their contextual dependencies across temporal-frequency dimensions. It operates through several principal phases: Firstly, a feature reconstruction module is designed to extract temporal, frequency, and temporal-frequency representations of emotional speech. Secondly, a high-low frequency-based emotion-aware partitioning strategy is designed to achieve the division of emotional regions. Thirdly, a local window within a hierarchical Transformer analyzes static intra-region correlations to capture fine-grained emotional patterns, while a dynamic window adaptively models temporal evolution across regions, learning dynamic inter-region relationships. Lastly, a dual-cross-attention mechanism is employed to synergize comprehensive emotion representation from different domains. Our evaluation experiments demonstrate that the proposed TF-DWFormer method achieves recognition accuracies of 73.68%, 91.67%, 92.59%, 74.42%, and 50.54% on the datasets IEMOCAP, CASIA, EMODB, eNTERFACE05, and MELD, respectively, outperforming existing SER methods. These results confirm the capability of TF-DWFormer to precisely localize salient regions, robustly model inter-region dependencies, and effectively fuse multi-domain information, providing a promising solution for advancing SER technology.}, archive = {J_EAAI}, author = {Yonghong Fan and Heming Huang and Huiyun Zhang and Ziqi Zhou}, doi = {10.1016/j.engappai.2025.112152}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112152}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Temporal-frequency joint hierarchical transformer with dynamic windows for speech emotion recognition}, volume = {161}, year = {2025}, }
• (2025). Three-dimensional reconstruction and fracture segmentation based on X-ray and computed tomography paired dataset. EAAI, 161, 112151. (www)

  @article{ , abstract = {In some orthopedic surgeries, the use of three-dimensional (3D) computed tomography (CT) scanning technology is not feasible due to scene limitations, leaving doctors to rely on two-dimensional (2D) X-ray images for real-time diagnosis. However, X-ray images lack 3D information, making accurate diagnosis challenging. Developing an algorithm to convert 2D X-ray images into 3D CT images, while simultaneously combining high-quality 3D reconstruction with precise fracture segmentation, offers a promising solution to the problem. In this study, we propose a novel artificial intelligence (AI)-driven framework named 3D reconstruction and segment anything model (3DRecSAM). The reconstruction image enhancer (RIE) is designed to achieve high-precision 3D reconstruction and provide high-quality feature initialization for fracture segmentation. Meanwhile, the mamba segment anything model (MSAM), based on the segment anything model (SAM) architecture, is developed for accurate fracture segmentation. We introduce a Kolmogorov–Arnold network (KAN)-based attention fusion module (KAF), which facilitates the joint optimization of the RIE reconstruction network and the MSAM segmentation network. Furthermore, the selective scanning mamba with KAN (SKM) is incorporated to enhance feature extraction for both RIE and MSAM. Mamba efficiently captures long-range dependencies and sequential patterns, while KAN’s learnable activation functions facilitate adaptive feature fusion and non-linear representation. To train and evaluate 3DRecSAM, we introduce the real X-ray and CT paired dataset (XCPData), which is publicly available on GitHub: https://github.com/YuanGao1201/XCPData .}, archive = {J_EAAI}, author = {Yuan Gao and Yuan Zhou and Da Chen and Jiachen Li and Mingle Zhou and Gang Li and Yunbo Gu and Jean-Louis Coatrieux and Yang Chen}, doi = {10.1016/j.engappai.2025.112151}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112151}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Three-dimensional reconstruction and fracture segmentation based on X-ray and computed tomography paired dataset}, volume = {161}, year = {2025}, }
• (2025). Unsupervised domain adaptation for lithology classification using dynamic entropy-based prototype learning. EAAI, 161, 112150. (www)

  @article{ , abstract = {Lithology classification plays a crucial role in geological exploration and resource evaluation. However, the significant distribution discrepancies between source and target domains, coupled with the unavailability of source domain data, pose substantial challenges to traditional domain adaptation methods. To address these challenges, we propose an innovative framework, Unsupervised Domain Adaptive Dynamic Entropy Prototype Learning (UDADEPL), which leverages a source-free unsupervised domain adaptation strategy for lithology classification. The UDADEPL framework consists of a frozen source pre-trained model and a trainable target model, incorporating a dynamic entropy-based prototype learning matrix for reliable sample selection and centroid-based pseudo-label learning for iterative optimization. Additionally, an information maximization loss and source domain regularization loss are integrated into a curriculum learning strategy to balance feature extraction and domain adaptation. This approach enables the model to effectively handle complex lithological boundaries and class imbalances in the target domain. Extensive experiments on datasets from the Tarim Oilfield and Hugoton–Panoma field demonstrate the superiority of UDADEPL over traditional machine learning and advanced deep learning models. UDADEPL achieves superior classification accuracy, outperforming the best baseline models, especially in cross-domain adaptation and complex lithology identification.}, archive = {J_EAAI}, author = {Hengxiao Li and Yahui Liu and Lu Liu}, doi = {10.1016/j.engappai.2025.112150}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112150}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unsupervised domain adaptation for lithology classification using dynamic entropy-based prototype learning}, volume = {161}, year = {2025}, }
• (2025). Transforming physics-informed machine learning to convex optimization. EAAI, 161, 112149. (www)

  @article{ , abstract = {Physics-Informed Machine Learning (PIML) offers a powerful paradigm of integrating data with physical laws to address important problems in engineering, such as parameter estimation, inferring hidden physics, equation discovery, and state prediction. However, PIML, such as Physics-Informed Neural Networks (PINNs), still lack the necessary accuracy, stability, and interpretability when applying in practical engineering due to many serious optimization challenges including the spectral bias, non-convex optimization, multi-objective optimization, and non-smooth optimization. In this study, we propose the Convex-PIML based on convex optimization and basis functions widely used in well-established numerical solvers to overcome all these limitations. The linear combination of B-splines is utilized to approximate the data, promoting the convexity of the loss function. By dividing variables into blocks and replacing the non-convex loss terms with convex approximations, the problem is further converted into a sequence of successively refined approximated convex optimization problems. This conversion known as Block Successive Convex Approximation (BSCA) allows the use of well-established convex optimization algorithms, obtaining solutions effectively and efficiently. Furthermore, an adaptive knot optimization method is introduced to mitigate the spectral bias issue of PIML, further improving the performance. The proposed fully adaptive framework by combining the adaptive knot optimization and BSCA is tested in scenarios with distinct types of physical prior. The results indicate that optimization problems are effectively solved in these scenarios, highlighting the potential of the framework for broad applications. Note that the Convex-PIML is also flexible since many other basis functions can also be incorporated to handle different systems.}, archive = {J_EAAI}, author = {Letian Yi and Siyuan Yang and Ying Cui and Zhilu Lai}, doi = {10.1016/j.engappai.2025.112149}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112149}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Transforming physics-informed machine learning to convex optimization}, volume = {161}, year = {2025}, }
• (2025). Finite-time fuzzy tracking control for networked suspension systems under gilbert-elliott fading channels. EAAI, 161, 112148. (www)

  @article{ , abstract = {In this paper, we address the finite-time ℓ 2 − ℓ ∞ fuzzy tracking control problem for networked suspension systems operating under Gilbert-Elliott fading channels. Firstly, the suspension system is established based on interval type-2 fuzzy rules to effectively describe the nonlinearity and uncertainty inherent in the model. Secondly, the Gilbert-Elliott model is employed to characterize the random switching behavior of channel states (good channel and fading channel) under real-world conditions. The timely identification of the current channel state is achieved through a mode detection method. To address channel fading and communication constraints, a novel mode-dependent controller is proposed. Simultaneously, we aim to minimize unnecessary and redundant communication. The hybrid scheduling strategy of the FlexRay protocol is utilized to transmit the measurement output data to the observer, and a saturation function is introduced in the observer to minimize the impact of outliers. Subsequently, the design conditions for stochastic finite-time boundedness and ℓ 2 − ℓ ∞ performance of the augmented system is proposed. Finally, simulation results demonstrate that the proposed algorithm is practically viable in dealing with complex environments, such as Gilbert-Elliott fading channels, outliers, and communication constraints.}, archive = {J_EAAI}, author = {Li-Juan Cai and Xiao-Heng Chang and Xin-Yue Zhao}, doi = {10.1016/j.engappai.2025.112148}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112148}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Finite-time fuzzy tracking control for networked suspension systems under gilbert-elliott fading channels}, volume = {161}, year = {2025}, }
• (2025). RT-FedFlow: An efficient framework for real-time traffic signal optimization using federated multi-agent reinforcement learning. EAAI, 161, 112147. (www)

  @article{ , abstract = {The optimization of Traffic Signal Control (TSC) is a critical component of the Intelligent Transportation System (ITS) in smart cities, aiming to eliminate traffic congestion and improve urban mobility. However, traditional approaches to TSC often rely on centralized data processing, limiting scalability and adaptability to real-time traffic dynamics while raising concerns about data privacy, communication, and computation. To address these challenges, this paper proposes the RT-FedFlow, a novel federated multi-agent reinforcement-learning (FMARL) framework for real-time traffic signal optimization. The framework employs enhanced FMARL techniques to enable decentralized collaboration among multiple traffic signal agents, optimizing traffic flow while seamlessly balancing local adaptability and global coordination. Federated model aggregation enables agents to collaboratively learn a global policy while preserving local data privacy, with each reinforcement learning (RL) agent independently optimizing its policy. Next, an advanced reward mechanism is designed to minimize congestion and reduce waiting times. Moreover, a hierarchical communication and coordination strategy is adopted to facilitate efficient local interaction among agents and global model aggregation to improve overall system performance. In addition, the framework enables real-time optimization and prioritization of emergency vehicles through a dedicated preemption module. Compared to existing TSC models, RT-FedFlow significantly improves traffic signal coordination and response times by dynamically adapting to complex, real-time traffic conditions. The proposed RT-FedFlow framework is implemented and evaluated using the SUMO simulator. Extensive simulation results demonstrate that RT-FedFlow outperforms all benchmark models, achieving a 15.49 % reduction in average inference time, a 3.07 % reduction in average queue length, a 13.04 % reduction in average intersection delay, a 0.20 % improvement in fuel consumption, and a 5.35 % emergency vehicle clearance time.}, archive = {J_EAAI}, author = {Ayaz Akbar and Syed Sajid Ullah and Abdul Malik and Saeed Mian Qaisar}, doi = {10.1016/j.engappai.2025.112147}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112147}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {RT-FedFlow: An efficient framework for real-time traffic signal optimization using federated multi-agent reinforcement learning}, volume = {161}, year = {2025}, }
• (2025). A hybrid data-driven model for state of health estimation of lithium-ion battery with capacity recovery. EAAI, 161, 112146. (www)

  @article{ , abstract = {The failure of lithium-ion batteries has attracted the attention of researchers. Monitoring the degradation process through sensor-based state-of-health (SOH) assessment enables early detection of anomalies and facilitates timely interventions to prevent catastrophic failures. During battery degradation, the phenomenon of capacity recovery makes it complicated to accurately evaluate the SOH. However, existing works usually neglect this phenomenon. In this paper, we propose a hybrid model focusing on capacity recovery. The model consists of four modules: a feature extraction module, a data decomposition module, a global trend assessment module, and a local fluctuation assessment module. Specifically, in the feature extraction module, a convolutional neural network extracts features from the input data that effectively characterize the capacity degradation process. The data decomposition module applies empirical mode decomposition (EMD) to separate local fluctuations caused by capacity recovery from the global degradation trend. These components are then input into their respective assessment modules. Additionally, the local fluctuation assessment module incorporates an attention mechanism that adaptively identifies the similarity between extracted features and health states, thereby enhancing the evaluation of local fluctuations. The proposed model is validated using the NASA (National Aeronautics and Space Administration) battery dataset, demonstrating its effectiveness in addressing the challenges of SOH prediction under capacity recovery conditions.}, archive = {J_EAAI}, author = {Yu Lin and Luo Zhou and Jianhai Yan and Shuguang He}, doi = {10.1016/j.engappai.2025.112146}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112146}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A hybrid data-driven model for state of health estimation of lithium-ion battery with capacity recovery}, volume = {161}, year = {2025}, }
• (2025). A novel localization algorithm integrated with sensor fusion for position estimation of mobile robot in the road following and roundabout environment. EAAI, 161, 112145. (www)

  @article{ , abstract = {A novel Artificial Intelligence algorithm based autonomous mobile robot localization in road environments is presented in this paper. The sensor fusion data, employing camera and odometry, are used to build the local maps for road environment. The localization of the robot is performed using a novel Artificial Intelligence algorithm, called Priori-Posteriori-Laser-Simulator, to find the position and heading-angle of robot along movement in two phases, namely priori and posteriori stages as 1st phase and Kalman filter in 2nd phase. In the priori stage, the path of the robot is planned in the built-map using Laser Simulator through generating a series of points as lines to drive the robot in the middle of the road environments. In the posteriori-stage, the position and robot heading angle are measured using sensor fusion system and the laser-simulator updates the position accordingly through a comparison between the planned and actual positions. The Kalman filter is used in 2nd phase to remove noise and enhance to enhance the localization accuracy. The proposed algorithm has been tested in indoor and outdoor road environments to show its effectiveness and performance for localizing robot in such environments. It is also compared with fuzzy-logic, adaptive H-infinity filter, fuzzy logic with Kalman filter and Oriented-Rotated Binary Simultaneous Localization and Mapping 2 algorithms on localizing robot at road environment. Results show the capability of the proposed algorithm to enable the robot to move effectively in road environments, recognize the features of road terrains and localize the robot during autonomous navigation in road environments.}, archive = {J_EAAI}, author = {Mohammed A.H. Ali and Nukman Yusoff and Bushroa Abd Razak and Sherzod Turaev and Rawad Abdulghafor and Aisha Muhammad}, doi = {10.1016/j.engappai.2025.112145}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112145}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel localization algorithm integrated with sensor fusion for position estimation of mobile robot in the road following and roundabout environment}, volume = {161}, year = {2025}, }
• (2025). Electroencephalography-based emotion recognition using a dual-stream multi-scale spatiotemporal convolutional capsule network. EAAI, 161, 112144. (www)

  @article{ , abstract = {In recent years, emotion recognition using electroencephalogram (EEG) signals has gained significant attention. However, shallow convolutional neural networks used in EEG emotion recognition are ineffective at capturing spatial relationships between features, adversely affecting model performance. To solve this problem, this study proposes a dual-stream multi-scale spatiotemporal convolutional capsule network aimed at improving EEG-based emotion recognition. The novelty of our approach is the dual-stream multi-scale spatiotemporal design, which enables parallel extraction and fusion of temporal and spatial EEG features at multiple scales, offering a richer and more discriminative representation for emotion recognition. Specifically, a dual-stream feature construction module is developed to extract multi-scale temporal and spatial features from raw EEG signals. A hybrid spatiotemporal attention mechanism enhances feature fusion, while a capsule-based classifier improves recognition accuracy by modeling relationships between local and global features. Experimental results using subject-dependent 10-fold cross-validation show average accuracies of 97.72%, 97.56%, and 97.82% for valence, arousal, and dominance on the Dataset for Emotion Analysis using Physiological signals(DEAP), with average F1 scores of 97.83%, 97.51%, and 97.68%. For the Dataset for Emotion Analysis using Physiological signals(DREAMER 1 ), the average accuracies are 96.48%, 96.32%, and 96.35%, with corresponding F1 scores of 96.23%, 95.74%, and 95.66%. The proposed method outperforms existing state-of-the-art approaches on both datasets, reducing the number of parameters by 58.07% and decreasing inference time by approximately 50.86% and 33.39%, compared to Residual Network 18. Additionally, in the subject-independent leave-one-subject-out cross-validation, the proposed method demonstrated results that significantly outperformed the baseline model across both datasets. Experiments show that the proposed method enhances spatiotemporal feature extraction and improves emotion recognition accuracy. This method reduces computational resource consumption and enhances recognition accuracy, thereby facilitating efficient algorithm development and deployment for real-time emotion monitoring and related applications.}, archive = {J_EAAI}, author = {Han Cai and Pengfei Lu and Xiaofang Wang and Yuxing Wang}, doi = {10.1016/j.engappai.2025.112144}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112144}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Electroencephalography-based emotion recognition using a dual-stream multi-scale spatiotemporal convolutional capsule network}, volume = {161}, year = {2025}, }
• (2025). Inverse structural design with generative and probabilistic autoencoders and diffusion models. EAAI, 161, 112143. (www)

  @article{ , abstract = {Traditional structural design is a forward trial-and-error process. Designers need to iterate through different design solutions and conduct structural analysis until the design meets the codes and standards. This study proposes and investigates a generative machine learning (ML) framework for inverse design of continuous beam systems. Three generative ML models, including conditional variational autoencoder (CVAE), conditional autoencoder with maximum likelihood estimation (CAE-MLE), and denoising diffusion models (DDMs) are trained and fine-tuned on the CBeamXP (Continuous Beam Cross-section Predictors) dataset with 1,000,000 beam sections to generate the cross sectional properties. Research results show that CAE-MLE achieves the highest generation accuracy and robustness, while CVAE offers more variability through latent space sampling. DDMs provide controllable generation variability via a stochasticity parameter in the inverse diffusion process. The proposed framework enables efficient generation of multiple design solutions and can potentially accelerate the conceptual design workflows in structural engineering. This work also demonstrated the feasibility toward artificial intelligence (AI)-assisted structural design using generative approaches and tabular datasets.}, archive = {J_EAAI}, author = {Bozhou Zhuang and Adrien Gallet and Danny Smyl}, doi = {10.1016/j.engappai.2025.112143}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112143}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Inverse structural design with generative and probabilistic autoencoders and diffusion models}, volume = {161}, year = {2025}, }
• (2025). Multi-objective generative design framework and realization for quasi-serial manipulator: Considering kinematic and dynamic performance. EAAI, 161, 112142. (www)

  @article{ , abstract = {This paper proposes a framework for optimizing the linkage mechanism of a quasi-serial manipulator for target tasks. The process is illustrated through a case study of a two-degree-of-freedom (2-DOF) linkage mechanism, which significantly influences the workspace of the quasi-serial manipulator. First, a diverse set of quasi-serial mechanisms is generated with workspaces that satisfy the target task and is converted into three-dimensional computer-aided design (3D CAD) models. Then, kinematic and dynamic analyses are conducted to compute workspace and payload torque labels for surrogate model training. Adaptive sampling is employed to identify an appropriate amount of training data for accurate prediction, while ensemble models are utilized to enhance robustness. After model training, a multi-objective optimization problem is formulated under the mechanical and dynamic constraints of the manipulator. The design objective is to recommend quasi-serial mechanisms with optimized kinematic (workspace) and dynamic (payload torque) performance that fulfill the target task. To explore the underlying physics of the Pareto solutions obtained via the Non-Dominated Sorting Genetic Algorithm (NSGA-II), various data mining techniques are applied to extract design rules that offer practical guidance. Finally, a detailed manipulator is realized using 3D-printed parts with topology optimization, and its performance is verified through a payload test. These results demonstrate the potential of the proposed framework for broader mechanism applications and its capability to support practical design decisions through design rule extraction.}, archive = {J_EAAI}, author = {Sumin Lee and Sunwoong Yang and Namwoo Kang}, doi = {10.1016/j.engappai.2025.112142}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112142}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-objective generative design framework and realization for quasi-serial manipulator: Considering kinematic and dynamic performance}, volume = {161}, year = {2025}, }
• (2025). Blockchain and deep learning-enabled IoT device-to-device authentication approach for smart cities using 5th generation technology. EAAI, 161, 112141. (www)

  @article{ , abstract = {As smart cities evolve, the increasing number of Internet of Things (IoT) devices requires secure authentication mechanisms for device-to-device (D2D) communication, especially within 5G networks. To address security challenges in D2D communication, Blockchain (BC) technology is utilized. Existing methods face issues like single-point failure attacks, high implementation costs, and data privacy concerns. This work aims to develop a secure and efficient authentication model leveraging BC technology and a Deep Q Network (DQN) for key generation to enhance IoT security in smart city applications. A novel authentication model, Deep Q Network-SecAuth (DQN-SecAuth), is proposed for secure D2D communication using BC. The model involves key entities such as the Registration Authority Center (RAC), blockchain, and IoT devices. The authentication process includes six stages: setup, registration, key generation, authentication, new device addition, and formal verification. In the key generation stage, the secret key is generated using a Deep Learning (DL) model named DQN. The proposed model also incorporates encryption, hash functions, Chebyshev polynomials, and XOR operations to ensure security. The DQN-SecAuth model achieved a minimum computational time of 9.525 s and memory usage of 42.897 megabytes (Mb), consensus delay of 0.420 s, energy consumption of 0.715 J, latency of 0.469 s, power consumption of 0.270 kW-hour (kWh), and throughput of 0.804 megabits per second (Mbps). The key novelty of this work is the use of a DQN for adaptive key generation, integrated into a BC-based authentication framework. This combination enhances security, reduces computational overhead, and outperforms traditional static key-generation and authentication methods in smart city IoT environments.}, archive = {J_EAAI}, author = {K. Prabhu Chandran and Bhuvaneswari P and Sivasankaran V and Vimala S}, doi = {10.1016/j.engappai.2025.112141}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112141}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Blockchain and deep learning-enabled IoT device-to-device authentication approach for smart cities using 5th generation technology}, volume = {161}, year = {2025}, }
• (2025). Emergency medical facility site selection in drone-based relief operations using an enhanced T-spherical fuzzy frank combined compromise solution method. EAAI, 161, 112140. (www)

  @article{ , abstract = {Unmanned aerial vehicles (UAVs) play a vital role in disaster response operations, emphasizing the need for efficient site selection for emergency medical facilities (EMFs). This study presents a structured framework for evaluating candidate sites and identifying critical criteria for establishing EMFs in drone-based relief operations. It proposes a novel T-spherical fuzzy (T-SF) multi-criteria group decision-making framework for EMF site selection. The framework integrates the T-SF-Entropy method and a hybrid subjective–objective weighting scheme that combines stepwise weight assessment ratio analysis and the method based on the removal effects of criteria. This integration enables a more reliable aggregation of expert opinions. The framework applies T-SF Frank aggregation operators in the modified combined compromise solution method to rank potential sites under multi-disaster conditions. A real-world case study evaluates alternative sites using the critical criteria and demonstrates the practical utility of the proposed model in the field of emergency engineering logistics. Quantitative analysis shows that the “urban logistics hub” achieves the highest compromise index due to its advantages in proximity, infrastructure, and supply chain access. Sensitivity analysis confirms that variations in parameters do not affect the stability of rankings. This study uses artificial intelligence to support intelligent decision-making in disaster response. It provides an effective engineering application that optimizes UAV-based healthcare deployment, enhances resource allocation, and improves the speed and accuracy of emergency medical responses.}, archive = {J_EAAI}, author = {Rajdip Mahajan and Saptadeep Biswas and Vladimir Simic and Dragan Pamucar and Abhijit Baidya and Uttam Kumar Bera}, doi = {10.1016/j.engappai.2025.112140}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112140}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Emergency medical facility site selection in drone-based relief operations using an enhanced T-spherical fuzzy frank combined compromise solution method}, volume = {161}, year = {2025}, }
• (2025). Feature description attention: Channel-independent local–global fusion for multi-scale feature representation. EAAI, 161, 112139. (www)

  @article{ , abstract = {In recent deep convolutional networks, attention mechanisms rely heavily on global information to generate attention weights, but few methods can effectively recalibrate global features based on local object characteristics. To address this gap, we introduce a masked-averaging strategy that adaptively selects regions of interest from the feature map, allowing local features to encode and reflect global information. By combining these local descriptors with global averages and maximums into a multi-scale Bag-of-Visual-Words (BoVW), our method jointly captures salient point, local region, and global context, resulting in richer and more discriminative feature representations. Additionally, we propose Feature Batch Normalization (FBN) to facilitate cross-channel interactions, further enhancing the performance of attention modules. In terms of engineering applications, to overcome the limitations of traditional channel-dependent attention under model compression and pruning, we design a channel-independent feature recalibration mechanism based on one-dimensional depthwise convolution, termed Feature Description Attention (FDA). It leverages the properties of BoVW and FBN to provide a flexible and lightweight attention module with minimal computational overhead and robust performance under quantization and pruning. Our method demonstrates consistent gains across three vision tasks: classification accuracy improves by 0.72–1.68%, object detection performance increases by 1.8–3.0% mean average precision, and segmentation accuracy rises by 1.9%. These results across different benchmarks confirm the effectiveness and generalizability of our approach. FDA offers a scalable, hardware-friendly solution for modern vision applications, especially valuable in resource-constrained settings such as mobile deployment, autonomous systems, and medical image analysis.}, archive = {J_EAAI}, author = {Yuanyang Zhu and Guangjie Han and Hongbo Zhu and Fan Zhang}, doi = {10.1016/j.engappai.2025.112139}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112139}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Feature description attention: Channel-independent local–global fusion for multi-scale feature representation}, volume = {161}, year = {2025}, }
• (2025). Improving robustness of transformers for power quality disturbance classification via optimized relevance maps. EAAI, 161, 112138. (www)

  @article{ , abstract = {Power quality disturbances (PQDs) are a critical area of research in the power systems domain. The growing complexity of modern power systems increases the variability of PQDs that may occur, making it challenging to manage power quality using traditional methods, and motivates the inclusion of machine learning models. However, these models are vulnerable to adversarial attacks. The predominant defense approach is adversarial training, which increases robustness at the price of significant increase in computational complexity of the training process, and can often lead to reduced accuracy on non-adversarial examples. In this light, this paper introduces a distinct approach to increase the robustness of Transformer-based PQD classifiers by leveraging their intrinsic explainability derived from the self-attention mechanism. By incorporating the resulting explainable features in the form of relevance maps into the loss function during the fine-tuning stage, we guide the model to focus on salient disturbance features. This method enhances the robustness against adversarial attacks, without compromising accuracy, and avoids the additional computational overhead associated with adversarial training. Experimental results show that when no adversarial attack is applied, the fine-tuning stage increases the PQD classification accuracy from 97.62% to 99.26%, thus overcoming the accuracy-robustness trade-off. Furthermore, the Transformer with the proposed fine-tuning stage is more adversarial robust than a state-of-the-art deep Convolutional neural network (DeepCNN) PQD classifier. The proposed Transformer maintains accuracy and F1 scores slightly above 88%, whereas fo DeepCNN both scores are slightly below 66.60%, thus highlighting the effectiveness of our approach.}, archive = {J_EAAI}, author = {Itamar Kapuza and Elinor Ginzburg-Ganz and Ram Machlev and Yoash Levron}, doi = {10.1016/j.engappai.2025.112138}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112138}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Improving robustness of transformers for power quality disturbance classification via optimized relevance maps}, volume = {161}, year = {2025}, }
• (2025). Progressive epipolar geometry for robust light field super-resolution. EAAI, 161, 112137. (www)

  @article{ , abstract = {Light field technology captures both spatial and angular information, densely sampled high-resolution light field images contain abundant 3-dimensional geometric information, enabling widespread applications in various industrial fields. However, existing methods for joint spatial and angular super-resolution of light field images face significant challenges when reconstructing scenes with large disparities and significant occlusions. To address this issue, we propose a geometric information enhancement module that enables efficient extraction and preservation of geometric information, and then we propose a non-disparity-based, one-stage approach that allows for the direct reconstruction of densely sampled high-resolution light field images from sparsely sampled low-resolution counterparts. Specifically, we decompose the original 4-dimensional light field image into three distinct 2-dimensional representations: spatial, angular, and epipolar plane image. Among these representations, the epipolar plane image contains abundant geometric information, which inspires us to propose a more efficient feature extractor. In addition, we introduce a progressive feature fusion strategy that better preserves geometric information extracted from epipolar plane images. Finally, to avoid errors introduced by warping operations that complicate the process of enhancing geometric information, we introduce a spatial-angular integrated upsampling module. Extensive experimental results on public datasets demonstrate that our proposed method significantly outperforms state-of-the-art approaches both quantitatively and qualitatively. Specifically, on the Occlusions dataset, our method achieved significant improvement in performance while reducing inference time by approximately 80% compared to the current best method. This efficiency gain is particularly beneficial for practical applications of the artificial intelligence algorithm.}, archive = {J_EAAI}, author = {Hao Zhang and Hao Sheng and Rongshan Chen and Da Yang and Ruixuan Cong and Zhenglong Cui and Xuefei Huang and Guanqun Su}, doi = {10.1016/j.engappai.2025.112137}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112137}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Progressive epipolar geometry for robust light field super-resolution}, volume = {161}, year = {2025}, }
• (2025). Exploring knee joint moment prediction models for landing task with various model architecture: Leveraging inertial measurement unit sensors data and subject-specific attributes. EAAI, 161, 112136. (www)

  @article{ , abstract = {This study presents a comparative evaluation of deep learning models for predicting Three-Dimensional (3D) knee joint moments during landing tasks using Inertial Measurement Unit (IMU) data. We assess a range of architectures, from simple feedforward networks to complex Three-Dimensional Convolutional Neural Networks (3D-CNN), to balance predictive accuracy with computational cost. The 3D-CNN model achieved the highest performance, with relative Root Mean Squared Error (rRMSE) values of 6.08 % for the knee abduction moment, 5.3 % for the external rotation moment, and 4.9 % for the flexion moment. We further explored the effect of adding subject-specific features, which yielded modest performance gains. While these features provided some improvement, model performance primarily depended on sensor data, reinforcing the need to integrate both biomechanical and Artificial Intelligence-driven insights. From an engineering application perspective, our findings guide researchers and clinicians in selecting the most suitable model and sensor placement based on both accuracy and computational complexity. Specifically, we highlight thigh-mounted IMU as the most effective placement for capturing landing dynamics. These insights support the practical deployment of knee joint moment prediction systems in real-world settings, such as gyms or training centers, where compact sensors and limited computational resources are essential.}, archive = {J_EAAI}, author = {Tommy Sugiarto and Yi-Jia Lin and Ya-Wen Tu and Hsiao-Liang Tsai and Lin-Fen Hsieh and Chi-Tien Sun and Patrik Kutilek and Wei-Chun Hsu}, doi = {10.1016/j.engappai.2025.112136}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112136}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Exploring knee joint moment prediction models for landing task with various model architecture: Leveraging inertial measurement unit sensors data and subject-specific attributes}, volume = {161}, year = {2025}, }
• (2025). Key region semantic information augmented transformer for image captioning. EAAI, 161, 112135. (www)

  @article{ , abstract = {Existing image captioning models often face difficulties in capturing inter-object relationships and generating description that comprehensively understands the entire image content, either relying on object detectors that overlook contextual information or depending on grid features that fail to adequately model spatial interactions. This paper proposes two solutions to these challenges. The first is the introduction of a module for mining semantic information from key regions. Based on the spatial proximity and high co-occurrence between objects, this module identifies the public region covered by these objects as a key region, mines their semantic information, and incorporates it into the modeling process, which compensates for the limitations of grid features. Second, we improve the standard Transformer decoder’s architecture by innovatively introducing an adaptive gating mechanism that dynamically adjusts the alignment between textual and visual features, enhancing the model’s overall comprehension of the image. To validate our approach, we applied these modules to the Transformer framework and proposed a novel method for image captioning, called Key region Semantic information Augmented Transformer (KSAT) for Image Captioning. Extensive experiments on benchmark datasets show that the proposed method outperforms many models. Specifically, our method achieves a score of 139.6% on the offline test, and 138.4% on the official online test server on the Consensus-based Image Description Evaluation (CIDEr) metric. In qualitative evaluation, our method also outperforms other methods at generating captions for complex scenes. Overall, these results confirm the validity of our method and advance the field of artificial intelligence.}, archive = {J_EAAI}, author = {Fuyun Deng and Wei Li and Zhixin Li}, doi = {10.1016/j.engappai.2025.112135}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112135}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Key region semantic information augmented transformer for image captioning}, volume = {161}, year = {2025}, }
• (2025). Enhanced neural-network-based iterative learning control considering iterative uncertainties for piezoelectric actuated micro-positioning platform. EAAI, 161, 112134. (www)

  @article{ , abstract = {This research aims to developing a new enhanced data-driven sliding-mode iterative learning control (E-DDSILC) strategy for piezoelectric actuated micro-positioning (PAMP) platforms. For the first time, the analysis demonstrating that errors converge to 0 in E-DDSILC is successfully extended from strictly repetitive systems to systems with non-strictly repetitive initial conditions. This generalization expands the practical application range of E-DDSILC. Simultaneously, iterative uncertainties are considered, which are the major factor affecting the performance of iterative learning control. To address these uncertainties, a diagonal recurrent neural network is employed to fit and compensate for them within a dynamic linearization model, thereby further enhancing the tracking accuracy and practicability of E-DDSILC. Finally, Several experiments are performed on a PAMP platform to compare the developed E-DDSILC method with both classical DDSILC and traditional E-DDSILC schemes. Comparative experimental results prove the superiority of the developed controller.}, archive = {J_EAAI}, author = {Miaolei Zhou and Yulong Sun and Xiuyu Zhang and Wei Gao and Chun-Yi Su}, doi = {10.1016/j.engappai.2025.112134}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112134}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhanced neural-network-based iterative learning control considering iterative uncertainties for piezoelectric actuated micro-positioning platform}, volume = {161}, year = {2025}, }
• (2025). A data-driven multi-stage stochastic optimization for sustainable humanitarian supply chain using machine learning algorithms. EAAI, 161, 112133. (www)

  @article{ , abstract = {The frequency and severity of natural disasters have intensified, resulting in significant human, financial, and emotional consequences. Earthquakes, in particular, have caused severe economic losses, deaths, and homelessness among millions. This study is designed to establish a comprehensive plan for managing pre- and post-disaster phases, including preparations, responses, and recovery efforts. It introduces a Multi-Stage Stochastic Programming (MSSP) model for sustainable humanitarian relief operations, optimizing location, allocation, and inventory management. The first and third stages concentrate on minimizing environmental impacts, while the second stage centers on enhancing social welfare. Simultaneously, economic cost reduction is consistently pursued in all three stages. The model's primary advantages include optimized inventory management to avoid shortages and flexible logistics strategies for timely and cost-effective delivery of relief items. Additionally, it ensures continuous aid, addressing both short-term and long-term needs to improve disaster management effectiveness and resilience. The multi-objective model is solved using Augmented Epsilon-Constraint (AEC). Furthermore, this paper employs Multi-Criteria Decision-Making (MCDM) methods to rank suppliers, leveraging Machine Learning (ML) algorithms to enhance ranking precision, thereby leading to a more responsive Supply Chain (SC). A real-world case study is then provided to illustrate the applicability and validity of the proposed model. Focusing on achieving balanced sustainability across all three stages, ensuring seamless logistics for all humanitarian supplies and affected individuals, and addressing uncertainties, the model determines the optimal quantities of all relief items to store. Through comprehensive analysis, the results provide key insights into the importance of MSSP in disaster management plans, enhancing understanding of the model's effectiveness.}, archive = {J_EAAI}, author = {Farnaz Ansari and Ali Bozorgi-Amiri and Hossein Shakibaei}, doi = {10.1016/j.engappai.2025.112133}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112133}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A data-driven multi-stage stochastic optimization for sustainable humanitarian supply chain using machine learning algorithms}, volume = {161}, year = {2025}, }
• (2025). A theme music generation model based on hybrid variational autoencoders and conditional generative adversarial networks. EAAI, 161, 112131. (www)

  @article{ , abstract = {The musical theme is the core melodic element that runs through a work. It is the main source of recognizable features and the basis of the overall structure of the score. However, existing music generation systems often face problems with missing thematic features and monotonous structure. These systems find it difficult to capture the complex thematic changes and long-term dependencies in music, resulting in a lack of dynamic changes, artistic expression in phrasing, and structural coherence in the generated results. In this paper, we propose a thematic music generation model based on a generative adversarial network, named Thematic Music Conditional Generative Adversarial Network (TM-CGAN), which solves these challenges through three key innovations. First, we propose a thematic fusion degree calculation module based on multi-dimensional feature metrics to overcome the theoretical limitations of existing methods, which lack explicit modeling of thematic features. Subsequently, we construct a two-dimensional image representation learning framework that preserves Musical Instrument Digital Interface (MIDI) symbolic sequences and employs convolutional neural networks (CNNs) to effectively model the multi-dimensional feature dependencies inherent in musical structures. Finally, we design a hybrid variational autoencoder-conditional generative adversarial network architecture that processes latent space modeling of thematic features through variational inference mechanisms while simultaneously optimizing generation quality through adversarial training. We conducted comprehensive experiments across three diverse MIDI music datasets to validate our approach. The experimental results demonstrate that TM-CGAN significantly outperforms state-of-the-art baseline models on multiple evaluation metrics, including generated music quality, theme representativeness, and structural integrity.}, archive = {J_EAAI}, author = {Fangzhu Jin and Peng Li and Xiaojun Wu}, doi = {10.1016/j.engappai.2025.112131}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112131}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A theme music generation model based on hybrid variational autoencoders and conditional generative adversarial networks}, volume = {161}, year = {2025}, }
• (2025). Malicious detection and trust calculation using residual recurrent neural network for trust with quality of service-aware multicast routing in mobile ad-hoc network system. EAAI, 161, 112130. (www)

  @article{ , abstract = {A Mobile Ad-hoc Network is a collection of mobile nodes without any proper infrastructure. In this work, a novel trust and Quality of Service-aware multicast routing technique for Mobile Ad-hoc Network is introduced. The key scope of this research paper is to evaluate the trust with Quality of Service -aware multicast routing process in Mobile Ad-hoc Network by detecting malicious nodes. For performing the optimal routing without any suspicious attacks initially the malicious node detection is performed by Residual Recurrent Neural Network. Further, if the Mobile Ad-hoc Network model is normal and free from malicious nodes, then the true value is calculated by utilizing the outcome of malicious node detection. If the node is free from malicious then, the trust value becomes high or else the trust value becomes low. Once, the trust calculation is completed the optimal routing is performed in the Mobile Ad-hoc Network with the support of the Enhanced Artificial Rabbits Optimization algorithm. Moreover, different constraints like hop count, throughput, Packet Delivery Ratio, delay, and energy consumption are derived. Further, different experiments are evaluated to prove the effectiveness of the implemented model against several baseline technique. Hence, the developed model accomplishes superior efficiency in detecting the malicious and performs the multicast routing in the Mobile Ad-hoc Network.}, archive = {J_EAAI}, author = {Sanjaya Kumar Sarangi and Rasmita Lenka and Janmejaya Mishra and Ritarani Sahu and Arabinda Nanda}, doi = {10.1016/j.engappai.2025.112130}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112130}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Malicious detection and trust calculation using residual recurrent neural network for trust with quality of service-aware multicast routing in mobile ad-hoc network system}, volume = {161}, year = {2025}, }
• (2025). Energy management system scheduling optimization based on an improved generative adversarial network deep reinforcement learning algorithm. EAAI, 161, 112129. (www)

  @article{ , abstract = {Households, as electricity consumers, play a critical role in achieving the carbon peak and carbon neutrality goals. The development of smart grids provides technical support for the efficient integration and distribution of renewable energy, gradually extending to household users. This has led to higher demands for the stability of electricity supply to address the growing electricity demand and the uncertainties associated with renewable energy. To address this, this paper proposes an improved generative adversarial network and an enhanced deep reinforcement learning algorithm to improve the scheduling capability of the energy management system. First, we introduce an improved wasserstein generative adversarial network that combines stochastic differential equations and autocorrelation penalty terms with the generator. The experimental results demonstrate that the proposed method can generate high-quality time series data. The generated data were used to train our scheduling model, effectively enhancing its generalization capability. Secondly, We introduced the Minmax mechanism to address Q-value estimation bias by utilizing multiple Q-networks. This mechanism first divides the target Q-values into several groups, selects the maximum value from each group, and then takes the minimum among these maxima as the final target Q-value. We applied this mechanism to improve deep reinforcement learning algorithms based on multi-Q-value evaluation. Comparison experiments show that this improvement significantly enhances the algorithm’s performance, outperforming traditional algorithms in terms of convergence, volatility, and final rewards. The energy management system demonstrates stronger adaptability when handling uncertainties arising from renewable energy variations, ensuring reliable power supply and achieving balanced energy management.}, archive = {J_EAAI}, author = {Weipeng Chao and Yuanbo Shi and Yushuai Li and Meng Liu and Xiaoling Leng}, doi = {10.1016/j.engappai.2025.112129}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112129}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Energy management system scheduling optimization based on an improved generative adversarial network deep reinforcement learning algorithm}, volume = {161}, year = {2025}, }
• (2025). Deep ensemble learning and error correction method for remaining useful life prediction of rolling bearings. EAAI, 161, 112128. (www)

  @article{ , abstract = {Rolling bearings are critical components of rotating machinery, and the prediction of their remaining useful life (RUL) is important for system reliability and operating efficiency. A novel RUL prediction method based on deep ensemble learning and error correction is here proposed. Firstly, the moving average filter (MAF) is applied to preprocess vibration signals for removing outliers and noise. Then, time-domain features are extracted from the processed vibration signals, and optimized to obtain imperative feature signals. Subsequently, a deep ensemble learning model is built with gated recurrent unit (GRU), bidirectional GRU (BiGRU), long short-term memory (LSTM) and bidirectional LSTM (BiLSTM) as base learners, and the overall performance of the prediction model is enhanced by introducing an error correction strategy. The MAF method is also used to smooth the trend of the RUL prediction outcomes. Finally, the proposed method is applied to two full-lifecycle rolling bearing datasets: the Prognostics and Health Management 2012 (PHM2012) dataset and the Intelligent Maintenance System (IMS) dataset. It is evaluated using mean square error (MSE), mean absolute error (MAE), and the R-square coefficient of determination (R 2 ). The test results demonstrate that the method achieves highly accurate RUL predictions: on the PHM2012 dataset, the MSE, MAE, and R 2 reach 0.000380, 0.013695, and 0.994716, respectively; on the IMS bearing dataset, the corresponding values are 0.001056, 0.015403, and 0.978346. In addition, the method outperforms traditional single models (GRU, BiGRU, LSTM, BiLSTM) as well as the Transformer model in both cases, with R 2 improvements over the Transformer of 0.011065 and 0.008744, respectively. These results highlight the superior generalization capability and robustness of the proposed method, making it applicable to industrial environments requiring reliable RUL prediction.}, archive = {J_EAAI}, author = {Wenzhe Yin and Hong Xia and Enrico Zio and Xueying Huang}, doi = {10.1016/j.engappai.2025.112128}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112128}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Deep ensemble learning and error correction method for remaining useful life prediction of rolling bearings}, volume = {161}, year = {2025}, }
• (2025). Maritime supply chain optimization using robust adversarial reinforcement learning. EAAI, 161, 112127. (www)

  @article{ , abstract = {This study explores novel strategies for analyzing and managing port productivity in a multi-stage supply chain network by integrating synchronization and reinforcement learning (RL) techniques. Current port management systems face issues with nonlinearity, high interdependency, and vulnerability to market disruptions, which might destabilize port operations. To tackle these issues, seaport operations are examined in four stages of implementation: the terminal operator, inland carrier, inland terminal operator, and consignee. Brownian motion is applied to characterize stochastic disruptions in volatile markets, and the port performance under market disruptions is analyzed using nonlinear data analytics. The dynamical analysis reveals that port management systems exhibit highly coupled nonlinear dynamics with a tendency towards instability. The complex nature of maritime port logistics requires innovative strategies to analyze container handling volumes, optimize the strategic planning of port management, and improve overall efficiency. A novel optimal policy for port operations is realized by integrating a deep deterministic policy gradient into robust adversarial deep learning. A deep reinforcement learning algorithm is employed to learn adaptively from historical port-related data and real-time container handling feedback, enabling intelligent strategies to make informed decisions and dynamically adjust policies in response to stochastic disruptions or changing market conditions. Quantitative results demonstrate that the proposed strategy achieves up to 97.78 % operating efficiency despite disturbances, and the adversarial attacks are shown to decrease port productivity by up to 77.78 % in scenarios without robust deep learning support. This study contributes to the growing field of intelligent port operations by paving the way for more adaptive and smart solutions in the maritime supply chain network.}, archive = {J_EAAI}, author = {Truong Ngoc Cuong and Sam-Sang You and Le Ngoc Bao Long and Hwan-Seong Kim and Duy Anh Nguyen and Nguyen Duy Tan}, doi = {10.1016/j.engappai.2025.112127}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112127}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Maritime supply chain optimization using robust adversarial reinforcement learning}, volume = {161}, year = {2025}, }
• (2025). Employing dual-path structure and soft attention mechanism to enhance recognition and classification of wild medicinal licorice in xinjiang. EAAI, 161, 112126. (www)

  @article{ , abstract = {Licorice is highly valued in traditional Chinese medicine for its anti-inflammatory, antiviral, and immunomodulatory properties, and is widely used in the pharmaceutical, food, and cosmetic industries. Xinjiang, the largest licorice-producing region in China, faces severe overharvesting of wild licorice due to increasing market demand, threatens natural populations and fragile ecosystems. Accurate identification and classification of licorice species are crucial for environmental protection and sustainable resource utilization, as traditional methods relying on experience are inefficient, subjective, and prone to errors. This study builds on the Inception-Residual Network-Version 2 (Inception-ResNet-V2) architecture and proposes an advanced licorice recognition model called Inception-ResNet-V2-Soft Attention, Dual-path Structure, and Focal Loss (IRV2-SDF). The IRV2-SDF model integrates a soft attention mechanism that focuses on key regions, a dual-path structure for multi-scale feature extraction, and a focal loss function to address class imbalance. It aims to improve the identification and classification of three wild licorice species ( Glycyrrhiza glabra , Glycyrrhiza inflata , and Glycyrrhiza uralensis ) and associated weeds in complex environments. Trained on 3,653 images collected from Xinjiang, the model achieves an average recognition accuracy of 91.79%, surpassing traditional models, with accuracy improvements of 4.27%, 2.08%, 2.76%, and 6.36% for G. glabra , G. inflata , G. uralensis , and weeds, respectively. By effectively reducing background noise and enhancing detection capabilities, the model overcomes the limitations of traditional methods and provides a robust solution for wild licorice recognition. This research offers a technical foundation for licorice conservation and sustainable utilization and can serve as a reference for identifying other medicinal plants in complex environments.}, archive = {J_EAAI}, author = {Yuan Qin and Jianguo Dai and Guoshun Zhang and Miaomiao Xu and Jing Yang and Jinglong Liu}, doi = {10.1016/j.engappai.2025.112126}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112126}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Employing dual-path structure and soft attention mechanism to enhance recognition and classification of wild medicinal licorice in xinjiang}, volume = {161}, year = {2025}, }
• (2025). Real-time cooperative target tracking in cluttered environments using multiple drone swarms with adaptive fuzzy emotional learning. EAAI, 161, 112125. (www)

  @article{ , abstract = {This paper presents a real-time adaptive trajectory prediction framework for cooperative unmanned aerial vehicle (UAV) swarms engaged in dynamic target tracking within cluttered environments. The proposed system introduces a novel artificial intelligence (AI)-based control architecture combining fuzzy inference, neuro-emotional learning, and distributed multi-agent coordination. At the core of the approach is a Bidirectional Fuzzy Brain Emotional Learning Prediction (BFBEL-P) model, which integrates fuzzy logic and an online adaptive neural structure to enable trajectory forecasting without pre-training or prior knowledge of the environment. From an engineering perspective, this AI model is deployed in UAV swarm navigation, where robust decision-making, predictive coordination, and obstacle avoidance are essential for target interception missions. In contrast to conventional prediction methods, such as curve fitting, nonlinear model predictive control (MPC), and deep learning-based Long Short-Term Memory (LSTM) networks, the BFBEL-P framework offers fast convergence, low computational cost, and high adaptability. The system incorporates multi-threaded data fusion across the swarm to achieve consensus-driven predictions and maintain situational awareness, even under sensor failures or occlusions. Simulation results show that BFBEL-P improves short-term prediction accuracy by 82.2%, reduces prediction time by 15%, and achieves a 100% tracking success rate across benchmark scenarios. These results establish BFBEL-P as a reliable AI technique for distributed control in real-world UAV swarm applications, offering a promising tool for search-and-rescue, surveillance, and defense operations.}, archive = {J_EAAI}, author = {Lucas William Page and Vu Phi Tran and Duy Luan Nguyen}, doi = {10.1016/j.engappai.2025.112125}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112125}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Real-time cooperative target tracking in cluttered environments using multiple drone swarms with adaptive fuzzy emotional learning}, volume = {161}, year = {2025}, }
• (2025). Hierarchical multi-scale matched masked autoencoder for industrial multi-rate time series modeling. EAAI, 161, 112122. (www)

  @article{ , abstract = {In practical industrial processes, due to sensor hardware limitations, the sampling rates of different variables often vary, leading to multi-rate time series (MRTS) data. However, the distribution of multi-scale dynamics in MRTS data typically follows a step-like pattern, with intricate scale transitions from fine to coarse and complex scale-consistent dependencies across rates. Additionally, the inherent characteristics of MRTS data often result in label scarcity. Both factors present significant challenges for MRTS modeling. To address these issues, we propose a novel self-supervised learning strategy, called H ierarchical M ulti-Scale M atched M asked A utoencoder (H3MAE). Specifically, we design a scale-matching input fusion mechanism where each layer is hierarchically aligned to a specific scale, with the scale-matching integration from two sources, effectively capturing the multi-scale dynamics and cross-rate scale-consistent dependencies in MRTS data. Besides, we introduce a novel auxiliary task that imputes masked positions in the encoded representation space at each layer, aiming to achieve MRTS representation learning and mitigate label scarcity. Furthermore, we propose a unique encoder-imputer structure in each layer to enable multi-scale self-supervised learning while generating temporally aligned features satisfying the input requirements of the next layer. Experimental results on three benchmark datasets and two industrial multi-rate tasks demonstrate that our framework yields better performance in MRTS modeling. The code is publicly available at https://github.com/monolithycq/H3MAE .}, archive = {J_EAAI}, author = {Changqing Yuan and Yongfang Xie and Shiwen Xie and Jie Wang}, doi = {10.1016/j.engappai.2025.112122}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112122}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Hierarchical multi-scale matched masked autoencoder for industrial multi-rate time series modeling}, volume = {161}, year = {2025}, }
• (2025). Unsupervised feature selection via latent feature representation and modified graph embedding. EAAI, 161, 112121. (www)

  @article{ , abstract = {Data in practical applications are correlated not only between samples but also between high-dimensional features. Latent representations can effectively represent such correlations. Existing latent representation methods mainly consider inter-instance relationships and rely on constructing similarity graphs. However, the interconnection information obtained by latent representations is redundant due to the noise and irrelevance of the raw data. Therefore, to address the above issues, this paper proposes an unsupervised feature selection by latent feature representation and modified graph (LFRMG, for short) embedding. First, a latent feature representation norm for learning self-representation structure is designed to explore the interconnections among features. Mining feature relationships via self-representation can solve the redundancy caused by fixed similarity graph relationships. Second, the latent feature representation is combined with modified graph regularization to achieve good feature interconnections while maintaining local data information. In addition, the l 2 , 0 -norm is utilized to select the full-row sparse projection of salient features to avoid the drawbacks of sparsity limitation and parameter tunability. A scheme for solving the closed form of l 2 , 0 is constructed. Finally, the presented approach is superior to many state-of-the-art unsupervised methods through comprehensive experiments on nine existing datasets.}, archive = {J_EAAI}, author = {Jialing Yan and Gang Hu and Guo Wei}, doi = {10.1016/j.engappai.2025.112121}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112121}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unsupervised feature selection via latent feature representation and modified graph embedding}, volume = {161}, year = {2025}, }
• (2025). Coarse-to-fine dual-branch network for ship target recognition in complex environments. EAAI, 161, 112120. (www)

  @article{ , abstract = {Harsh sea conditions and the complex and variable positions of ships significantly impact the capacity of imaging devices to capture high-quality ship images, making ship target recognition challenging in the application of artificial intelligence. Many scholars have recently proposed cascaded recognition models to address this issue. Following this method, in this paper, we propose a novel method for ship target recognition in complex environments called the coarse-to-fine dual-branch (CFDB) network. The CFDB model designs a dual-branch network from coarse to fine to lock the target area fine features and then uses peer-to-peer communication to extract and exchange learning of the target region’s final discriminative contour features, assisting in predicting ship classes in the complex environment. The proposed method is evaluated on the constructed complex in background ships (CIB-ships) dataset and the publicly available Marine Argos Recognition Ships (MAR-ships) and Game-of-Ships datasets. Compared with the suboptimal method, the proposed CFDB network exhibits improvements of 2.11%, 1.33%, and 1.24% accuracy on the CIB-ships, MAR-ships, and Game-of-ships datasets, respectively. The results demonstrate that the proposed method provides useful ideas for the dynamic monitoring of ships in real environments. Our code will be published at https://github.com/yangt1013/CFDB-master .}, archive = {J_EAAI}, author = {Yang Tian and Hao Meng}, doi = {10.1016/j.engappai.2025.112120}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112120}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Coarse-to-fine dual-branch network for ship target recognition in complex environments}, volume = {161}, year = {2025}, }
• (2025). A multi-scale sparse channel transformer network for image reconstruction of astronomical bright source contamination. EAAI, 161, 112119. (www)

  @article{ , abstract = {Bright source contamination has long been a challenging issue in the field of image processing, particularly in applications such as astronomical observations, satellite imaging, and nighttime surveillance. To address this issue, this paper proposes a novel Multi-Scale Sparse Channel Transformer Network (MSCformer) aimed at achieving high-quality image reconstruction under the influence of bright source contamination. The network integrates a Top-k Sparse Attention mechanism with a Channel Attention module, enabling selective focus on the most informative features and adaptive weight allocation across channels. Additionally, a Multi-Scale Dual-Gate Feedforward Network is designed to further enhance the expression of valuable features while suppressing redundant information. Experimental results demonstrate that the proposed method exhibits outstanding performance in practical applications on the Sloan Digital Sky Survey (SDSS) photometric image dataset. Compared to existing state-of-the-art techniques, MSCformer achieves significant performance improvements, with a Peak Signal-to-Noise Ratio (PSNR) of 45.093 decibel(dB), a Structural Similarity Index Measure(SSIM) of 0.978, and a Pixel Average Absolute Error (PAAE) of 0.675. This not only significantly enhances the removal of bright source contamination in the field of astronomy but also provides important reference value for subsequent research in related domains.}, archive = {J_EAAI}, author = {Yajuan Zhang and Congcong Shen and Xia Jiang and Bo Qiu and Ali Luo and Fuji Ren and Yuanlu Chen}, doi = {10.1016/j.engappai.2025.112119}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112119}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-scale sparse channel transformer network for image reconstruction of astronomical bright source contamination}, volume = {161}, year = {2025}, }
• (2025). Seismic damage assessment of reinforced concrete frame structures based on transfer-residual networks. EAAI, 161, 112118. (www)

  @article{ , abstract = {Post-earthquake building damage assessment is a critical research focus in civil engineering due to its vital role in guiding timely and informed rescue operations. This study proposes a rapid seismic damage assessment method for reinforced concrete frame structures based on transfer-residual networks. The methodology involves pre-training a base model on a large dataset generated from computationally efficient simplified model simulations, followed by transfer learning on a smaller, high-fidelity dataset derived from refined finite element model simulations. This approach significantly enhances prediction accuracy while reducing computational costs for seismic damage assessment in new building structures. Optimization of model parameters was performed to find the optimal residual network, which achieves an accuracy of 91.2 % with a remarkable 32-fold speedup over conventional nonlinear time-history analysis methods. Moreover, the transfer-residual network enhances accuracy by 10 % compared to training from scratch. The results conclusively demonstrate the feasibility of utilizing large datasets from simplified models to improve the training accuracy of refined models with limited data, providing a valuable reference for advancing transfer learning-based frameworks in seismic damage assessment of buildings.}, archive = {J_EAAI}, author = {Chen Xiong and Zhijie Luo and Jie Zheng and Linlin Xie and Liu Mei and Lixiao Li and Wujian Long}, doi = {10.1016/j.engappai.2025.112118}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112118}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Seismic damage assessment of reinforced concrete frame structures based on transfer-residual networks}, volume = {161}, year = {2025}, }
• (2025). Machine learning-enabled performance-based design of three-dimensional printed engineered cementitious composites. EAAI, 161, 112117. (www)

  @article{ , abstract = {The superior tensile ductility of engineered cementitious composites (ECC) offers a promising solution to the challenge of integrating conventional steel reinforcement in three-dimensional (3D) concrete printing (3DCP). However, the widespread adoption of 3D printed ECC (3DP-ECC) is hindered by the reliance on trial-and-error design process. The complex material component and inherent anisotropy of 3DP-ECC pose challenges for accurate property prediction and inverse design. This paper introduces a performance-based design strategy for 3DP-ECC, leveraging machine learning (ML) and multi-objective optimization. The anisotropic-mechanical properties including compressive strength and flexural strength were experimentally and statistically investigated; further, ML prediction models conbined with multi-objective optimization algorithm were developed to inversely design 3DP-ECC for specific mechanical performance requirements, while reducing carbon footprint and material cost. Specifically, an extensive database was assembled, followed by grey relational analysis (GRA) to identify the parametric sensitivity of the mechanical properties of 3DP-ECC. Three representative ML techniques were employed, with the back-propagation artificial neural network (BPANN) demonstrating superior predictive accuracy. Model interpretability analyses uncovered the importance of input parameters and their influence on predicted outcomes. Lastly, non-dominated Sorting Genetic Algorithm II (NSGA-II) integrated with the BPANN models was applied to perform the inverse design of 3DP-ECC, showing good effectiveness and accuracy. This work offers an efficient and viable avenue for performance-based design for 3DP-ECC, along with the potential to develop low-carbon cost-effective 3DP-ECC.}, archive = {J_EAAI}, author = {Wenguang Chen and Long Liang and Junhong Ye and Lingfei Liu and Neven Ukrainczyk and Liqiang Yin and Jiangtao Yu and Kequan Yu}, doi = {10.1016/j.engappai.2025.112117}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112117}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Machine learning-enabled performance-based design of three-dimensional printed engineered cementitious composites}, volume = {161}, year = {2025}, }
• (2025). Personalized text-to-image generation with large language and vision assistant enhanced training. EAAI, 161, 112116. (www)

  @article{ , abstract = {Personalized image generation aims to synthesize images of a specific identity. The identity, denoted as V ∗ , refers to an entity with distinctive visual attributes, such as a dog-shaped backpack. However, existing methods like DreamBooth and Custom Diffusion often struggle to generate images of V ∗ that accurately match the input prompts. In this work, we analyze two key issues underlying this limitation: (1) the overbinding problem , where the prompt tokens used to represent V ∗ unintentionally bind to irrelevant visual details from the reference image during training; and (2) the low language prior problem , where insufficient use of pre-trained language prior limits the model’s ability to faithfully generate all the prompt words. To overcome these challenges, we propose LLaVA-Booth, a novel personalization method for diverse, identity-preserving image generation, based on Large Language and Vision Assistant (LLaVA) enhanced training. Our method alleviates the overbinding problem by disentangling background information and solves the low language prior problem by enriching the language context. Additionally, we introduce two auxiliary objectives: (1) an identity (ID) binding loss to strengthen the identity binding and (2) a prior preservation loss to prevent language drift and encourage generation diversity. Experiments demonstrate that LLaVA-Booth effectively mitigates overbinding and enhances language priors to improve prompt fidelity, then generates diverse, high-quality, and identity-preserving images of V ∗ .}, archive = {J_EAAI}, author = {Junsheng Luan and Zhanjie Zhang and Wei Xing and Lei Zhao}, doi = {10.1016/j.engappai.2025.112116}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112116}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Personalized text-to-image generation with large language and vision assistant enhanced training}, volume = {161}, year = {2025}, }
• (2025). A multi-view new energy vehicle form generation design method combining kansei imagery and deep learning. EAAI, 161, 112115. (www)

  @article{ , abstract = {In the competitive landscape of new energy vehicles, exterior design has become a crucial differentiator amid functional homogenization. User preferences are central to shaping vehicle appearance, yet most perceptual design methods rely on a single viewpoint, limiting insights into complex preference patterns. This study proposes a multi-perspective mapping approach that integrates Kansei engineering with deep learning. Firstly, user core imagery is collected and mined through big data. Secondly, Kernels Network (KNet) semantic segmentation model, Residual Networks (ResNet) tri-view (front/side/rear) score prediction model and fully connected network (FCN) feature fusion model are integrated to construct a multi-view feature mapping system. Finally, the optimal combination of morphological elements is explored based on the Elite Genetic Algorithm (EGA), and the scheme is validated through generative artificial intelligence (AI) workflow. The experimental results demonstrate that, employing “Cool” as a case study, the three-view scheme and the combination scheme devised by this research process exhibit substantial superiority over the majority of the samples. Under identical parameters, the scheme with decision constraints surpasses the randomly generated scheme in terms of perceptual scores and stability. The performance of the test set and the experimental results collectively substantiate the model’s validity. This workflow—covering preference extraction, morphological decomposition, AI-driven generation, and validation—provides a scalable framework for new energy vehicle exterior design. It also demonstrates novel applications of Kansei engineering in multi-view fusion and generative form design.}, archive = {J_EAAI}, author = {Zihao Wang and Le Xi and Yifan Ding and Wenjie Fang and Kaiming Wang and Hongliang Zuo}, doi = {10.1016/j.engappai.2025.112115}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112115}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A multi-view new energy vehicle form generation design method combining kansei imagery and deep learning}, volume = {161}, year = {2025}, }
• (2025). Explainable remaining useful life uncertainty prediction method for rolling bearing. EAAI, 161, 112114. (www)

  @article{ , abstract = {Rolling bearing remaining useful life prediction is the core technology of equipment maintenance. Although deep learning-based prediction methods have made significant breakthroughs, the problems of insufficient model explainability and prediction result uncertainty quantification have seriously constrained the credibility of maintenance decisions. Therefore, this research combines prediction uncertainty quantification with model explanation to propose an explainable uncertainty prediction method. The method includes multi-dimensional feature extraction, remaining useful life uncertainty prediction, and Shapley additive explanations interpreter. For feature extraction, multi-dimensional feature vectors are constructed as network inputs by extracting time-domain features and frequency-domain features. Then, the remaining useful life prediction interval for the rolling bearing is compressed by the proposed gated temporal quantile network. Finally, the prediction model is explained using the Shapley additive explanations interpreter. The multi-case validation results based on the Xi'an Jiaotong University and Changxing Sumyoung Technology Co., Ltd. (XJTU-SY) and Intelligent Maintenance Systems (IMS) rolling bearing full life cycle datasets show that the proposed model has an average interval coverage of 93.96 % and an average interval width of 9.92 %, which indicates that the model maintains high accuracy and robustness in different cases. The nonlinear mapping relationship between the prediction results and the features is clarified by analyzing the Shapley values. Finally, the Shapley values are used to rank the importance of the features to locate the position that may cause rolling bearing performance degradation, which provides credible decision support for the development of the predictive maintenance strategy.}, archive = {J_EAAI}, author = {Ting Zhang and Honglei Wang}, doi = {10.1016/j.engappai.2025.112114}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112114}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Explainable remaining useful life uncertainty prediction method for rolling bearing}, volume = {161}, year = {2025}, }
• (2025). Machine learning-based estimation of sound transmission loss in single and double-layered rectangular functionally graded plates. EAAI, 161, 112112. (www)

  @article{ , abstract = {This study presents a robust machine learning (ML) framework for predicting sound transmission loss (STL) in temperature-dependent functionally graded (FG) single- and double-layered plates, combining analytical modeling and data-driven approaches. The FG plates are modeled with material properties varying along the thickness via a power-law distribution, incorporating nonlinear temperature effects. A high-fidelity analytical solution is derived using first-order shear deformation theory (FSDT) and acoustic velocity potential, with governing equations formulated via Hamilton's principle and solved using the weighted residual method. To enable ML training, a large-scale parametric study generates 189,000 data points, covering variations in plate geometry, temperature, acoustic cavity depth, gradient index, and incident wave angles. ML algorithms are trained on this data to predict STL, with the extreme gradient boosting (XGBoost) algorithm demonstrating the highest accuracy (coefficient of determination R 2 = 99.38 % for training data and R 2 = 99.11 % for test data). The validated ML model is then employed to investigate key parameters—temperature, power-law index, incidence angle, cavity depth, and plate thickness—revealing their nonlinear interactions and impact on STL performance. The developed artificial intelligence (AI) framework provides an efficient tool for acoustic design optimization, offering real-time parameter tuning capabilities for engineers working with functionally graded acoustic barriers.}, archive = {J_EAAI}, author = {Chunfeng Jiang and Masoud Babaei}, doi = {10.1016/j.engappai.2025.112112}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112112}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Machine learning-based estimation of sound transmission loss in single and double-layered rectangular functionally graded plates}, volume = {161}, year = {2025}, }
• (2025). Cubic linguistic T-spherical fuzzy aggregation operator-based multi-attribute group decision-making model and its application to food waste treatment technique selection. EAAI, 161, 112111. (www)

  @article{ , abstract = {Food waste presents significant environmental, economic, and social challenges globally, contributing to resource depletion and greenhouse gas emissions. Effective treatment techniques are essential for minimizing these impacts, promoting sustainability, and supporting circular economy practices. However, selecting the most suitable food waste treatment technique is a complex multi-attribute group decision-making (MAGDM) problem that involves the simultaneous consideration of environmental, economic, and social factors under significant uncertainty. To address this challenge, this study proposes a novel cubic linguistic T-spherical fuzzy sets (CLT-SFSs) framework, which integrates linguistic T-spherical fuzzy sets and linguistic interval-valued T-spherical fuzzy sets to enhance decision-making precision under uncertainty. The proposed CLT-SF framework offers a unified structure capable of capturing both qualitative expert opinions and quantitative uncertainty ranges, offering unprecedented expressiveness for complex sustainability assessments. First, the formal definition and basic operations for CLT-SF numbers, including addition, multiplication, scalar multiplication, and scalar power, and a comparison law are established. Next, to efficiently aggregate cubic linguistic T-spherical fuzzy information, we propose the cubic linguistic T-spherical fuzzy weighted averaging and the cubic linguistic T-spherical fuzzy weighted geometric aggregation operators. These aggregation operators can effectively and comprehensively aggregate attribute values in MAGDM problems. Subsequently, utilizing these operators, a CLT-SFS-based combinative distance-based assessment model is developed to address MAGDM problems. The model's applicability and robustness are demonstrated through a real-world case study on the selection of food waste treatment techniques. A parameter analysis is also conducted to examine the sensitivity of ranking outcomes. Finally, a comparative analysis with existing methods confirms the proposed model's effectiveness, feasibility, and advantages in addressing complex decision-making scenarios. This research not only advances the theoretical framework of fuzzy decision-making but also provides a practical tool for stakeholders in waste management to make informed and sustainable choices.}, archive = {J_EAAI}, author = {Shahid Hussain Gurmani and Weiping Ding and Rana Muhammad Zulqarnain and Jiangfeng Hao}, doi = {10.1016/j.engappai.2025.112111}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112111}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Cubic linguistic T-spherical fuzzy aggregation operator-based multi-attribute group decision-making model and its application to food waste treatment technique selection}, volume = {161}, year = {2025}, }
• (2025). Contextual and orientation correction modules enhance weakly-supervised aerial object detection in remote sensing images. EAAI, 161, 112110. (www)

  @article{ , abstract = {This paper addresses the challenges of context and orientation ambiguity in weakly supervised aerial object detection. While current research focuses on improving detection accuracy and efficiency, it often encounters difficulties with contextual and rotational variations in aerial imagery. We propose a novel Context and Orientation Correction (COC) framework, which includes two innovative modules: a context correction module and an orientation correction module. The context correction module utilizes style normalization to guide the model in identifying atypical objects within specific contextual scenes by mitigating contextual disparities between instances and refining contextual information. Additionally, the orientation correction module aims to reduce feature distance between instances with varying orientations, leveraging contrastive learning to ensure consistent object representations. Furthermore, we introduce a category-aware aggregation loss to enhance similarity in feature representations of objects from the same category, thereby addressing the class collision issue commonly associated with contrastive learning. Our COC framework achieves 27.6% mean Average Precision and 59.8% mean Average Precision on the Detection in Optical Remote Sensing Image (DIOR) and Northwestern Polytechnical University Very High Resolution 10. v2 (NWPU VHR-10.v2) datasets, respectively, demonstrating its significant effectiveness.}, archive = {J_EAAI}, author = {Le Yang and Shunzhou Wang and Xuerong Wang and Shutong Wang and Yuting Lu and Binglu Wang}, doi = {10.1016/j.engappai.2025.112110}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112110}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Contextual and orientation correction modules enhance weakly-supervised aerial object detection in remote sensing images}, volume = {161}, year = {2025}, }
• (2025). Towards detection of anomalous cosmic ray signals for observations acquired from cosmic ray extremely distributed observatory mobile detectors. EAAI, 161, 112109. (www)

  @article{ , abstract = {In this work, we propose and test a method that allows the detection of anomalous cosmic ray signals acquired using Complementary Metal-Oxide-Semiconductor detectors. The method uses unsupervised embedding based on Principal Component Analysis which we named Eigenhits in apparent analogy to Eigenfaces. The embedding generated using Eigenhits allows the detection of potential anomalies, defined as images whose position described by the embedding relative to a given measure is above a certain distance threshold from other images. Thus, the problem of anomaly detection was reduced to the problem of detecting outliers which can be solved, for example, using clustering algorithms. We conducted tests of our approach on the Cosmic Ray Extremely Distributed Observatory (CREDO) dataset containing 13168 images and obtained satisfactory results demonstrating the stability and effectiveness of the method. The embedding generation method we propose in this paper and the evaluation of its effectiveness in detecting anomalies in images of cosmic ray events from CREDO dataset is a pioneering study with many critical applications.}, archive = {J_EAAI}, author = {Tomasz Hachaj and Łukasz Bibrzycki and Marcin Piekarczyk and Olaf Bar and Michał Niedźwiecki and Sławomir Stuglik and Piotr Homola and Dmitriy Beznosko and David Alvarez-Castillo and Bożena Poncyljusz and Ophir Ruimi and Oleksandr Sushchov and Krzysztof Rzecki and CREDO collaboration}, doi = {10.1016/j.engappai.2025.112109}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112109}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Towards detection of anomalous cosmic ray signals for observations acquired from cosmic ray extremely distributed observatory mobile detectors}, volume = {161}, year = {2025}, }
• (2025). An uncertainty-aware safe-evolving reinforcement learning algorithm for decision-making and control in highway autonomous driving. EAAI, 161, 112108. (www)

  @article{ , abstract = {Rule-based and optimization-based approaches face challenges in decision-making and control for autonomous vehicles (AVs) in dynamic and complex highway scenarios. In contrast, reinforcement learning (RL) offers a more flexible and adaptable data-driven solution by allowing AVs to learn optimal actions through interactions with the environment, without requiring predefined rules or explicit programming. However, in real-world highway environments characterized by uncertainty, RL algorithm encounter difficulties in ensuring stability and safety. To address these challenges, this paper proposes an uncertainty-aware safe-evolving RL algorithm that integrates internal stability, external stability, and provable safety mechanisms. The internal stability mechanism ensures consistent performance improvements with high probability during policy updates itself, while the external stability leverages a benchmark policy as a reference to ensure the current policy performs at least as well as, if not better than, the benchmark. Furthermore, an action projection mechanism and a mixed learning procedure are incorporated to make minimal modifications to the learned policy, ensuring safety while supporting stable learning from both safe and original actions. The results show that the proposed algorithm maintains stability and safety throughout the learning process, achieves final performance comparable to traditional RL methods, and delivers higher training efficiency in a complex dynamic highway scenario in simulation. This suggests that the algorithm offers a viable solution for self-evolving systems in uncertain real-world environments, where traditional approaches may struggle.}, archive = {J_EAAI}, author = {Ping Lu and Sunan Zhang and Feihong Tan and Fulin Zhang and Yuxiang Feng and Bo Hu}, doi = {10.1016/j.engappai.2025.112108}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112108}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An uncertainty-aware safe-evolving reinforcement learning algorithm for decision-making and control in highway autonomous driving}, volume = {161}, year = {2025}, }
• (2025). A deep learning-driven network for joint low-light enhancement and deblurring in maritime surveillance systems. EAAI, 161, 112107. (www)

  @article{ , abstract = {—Shipboard video surveillance systems are crucial for maritime environmental perception. However, shipborne cameras are often affected by challenges such as ship rolling and low illumination, causing motion blur and low-light degradation in captured images. These degradations significantly reduce the accuracy and robustness of sea-surface obstacle detection. To address this challenge, this paper presents a deep learning-based joint network model designed to simultaneously deblur and enhance low-light maritime images. Unlike existing step-by-step approaches that treat deblurring and low-light enhancement as separate tasks, the proposed model employs an encoder-decoder architecture to jointly address both tasks within a unified framework, thereby overcoming the limitations of separate processing. Furthermore, the model incorporates a novel adaptive wavelet curve attention mechanism and an extended spatial pyramid pooling module with dynamic global context to mitigate detail loss caused by sea surface reflections and artifacts in dynamic backgrounds. Additionally, a joint loss function is exploited to further enhance image restoration quality. Experiments on a self-constructed synthetic maritime dataset demonstrate that the proposed model achieves a Peak Signal-to-Noise Ratio (PSNR) of 31.81, a Structural Similarity Index Measure (SSIM) of 0.812, and a Learned Perceptual Image Patch Similarity (LPIPS) of 0.246, significantly outperforming existing competing methods. Moreover, evaluations on real-world degraded maritime images confirm the model's strong generalization ability and robustness. This advancement offers practical benefits for improving obstacle perception in autonomous ships navigating complex maritime environments.}, archive = {J_EAAI}, author = {Zeyuan Shao and Yong Yin and Hongguang Lyu and Qianfeng Jing and Tao Cheng}, doi = {10.1016/j.engappai.2025.112107}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112107}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A deep learning-driven network for joint low-light enhancement and deblurring in maritime surveillance systems}, volume = {161}, year = {2025}, }
• (2025). A double hierarchy hesitant fuzzy forecasting model considering the influence of investor emotion and the co-movement of stock markets. EAAI, 161, 112106. (www)

  @article{ , abstract = {The stock market is a complex system influenced by various internal and external factors. Effective stock index forecasting models must balance interpretability with the ability to integrate multi-source information. While traditional fuzzy models offer high interpretability, they struggle to incorporate multiple dynamic factors. This study proposes a novel forecasting model based on double hierarchy hesitant fuzzy linguistic term sets (DHHFLTS), which integrates investor sentiment and international market co-movement. By constructing fuzzy logic rules, the model captures complex interactions underlying index fluctuations and enhances predictive accuracy in volatile environments. To evaluate its effectiveness, the model was tested on time series data from three major indices: the Shanghai Stock Exchange Composite Index (SSEC), Taiwan Capitalization Weighted Stock Index (TAIEX), and Financial Times Stock Exchange 100 Index (FTSE 100). The Root Mean Square Error (RMSE) achieved was 32.06 for the SSEC, 69.74 for the TAIEX, and 25.27 for the FTSE 100. Compared with existing benchmark methods, the proposed model demonstrates superior predictive accuracy and generalization performance. These findings confirm the model's capability in capturing multidimensional uncertainty and temporal patterns, highlighting its potential for intelligent financial forecasting and decision-making.}, archive = {J_EAAI}, author = {Liu Zhengmin and Du Chuantao and Zhang Jihao and Gao Shanshan and Liu Peide}, doi = {10.1016/j.engappai.2025.112106}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112106}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A double hierarchy hesitant fuzzy forecasting model considering the influence of investor emotion and the co-movement of stock markets}, volume = {161}, year = {2025}, }
• (2025). Style prompt tuning for bridging visual gaps in autonomous driving. EAAI, 161, 112105. (www)

  @article{ , abstract = {Artificial intelligence models for semantic segmentation and image classification in autonomous driving must maintain reliability across adverse conditions such as rain, fog, snow, and nighttime. However, models trained on only clear daytime images often fail to generalize under such domain shifts. Existing unsupervised domain adaptation (UDA) methods employ image-level style transfer using generative adversarial networks (GANs) or diffusion models, which necessitate paired data and risk altering content. Therefore, this study proposes Style Prompt Tuning, a novel UDA framework that utilizes image–text models to automatically generate and optimize textual prompts representing target-domain styles. These prompts guide a U-Net-based style network to synthesize source images in the target style while preserving their semantic content. Our approach employs clustering within the Contrastive Language—Image Pretraining (CLIP) embedding space and a composite loss function, including content, style, transfer, patch, and total variation terms to optimize prompt quality. The generated stylized images augment the source dataset and are used to train more robust task models. Experiments on semantic segmentation benchmarks (Cityscapes-to-Adverse Conditions Dataset with Correspondences (ACDC), DarkZurich, BDD100k-night, and Nighttime Driving) and image classification (Visual Domain Adaptation 2017) reveal our approach to achieve improvements of +3.4 mean intersection-over-union (mIoU) and +0.8% accuracy over prior UDA methods. These results highlight our method’s practical effectiveness for real-world autonomous driving applications under visually challenging scenarios.}, archive = {J_EAAI}, author = {Suyeon Cha and Giyun Choi and Minji Kwak and Jongwon Choi}, doi = {10.1016/j.engappai.2025.112105}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112105}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Style prompt tuning for bridging visual gaps in autonomous driving}, volume = {161}, year = {2025}, }
• (2025). Enhancing fuzzy cognitive map convergence through supervised and unsupervised learning algorithms: A case study of operational risk assessment in power distribution networks. EAAI, 161, 112104. (www)

  @article{ , abstract = {Fuzzy Cognitive Maps (FCMs) are commonly used for modeling complex systems. However, their convergence challenges significantly limit their accuracy and reliability, especially in operational risk assessment. To address this issue, the current study proposes a novel approach that integrates advanced supervised and unsupervised learning algorithms: specifically, the Mesh Adaptive Direct Search (MADS) and Genetic Algorithm (GA). To meet the critical need for accurate risk modeling in power distribution networks, the proposed methodology utilizes ten years of time-series data from the Yazd Power Distribution Network as a real-life case study. This optimizes risk relationships and reduces convergence errors. The main contributions of this research are: (1) the development of an integrated FCM-based framework that improves convergence stability and accuracy through advanced learning algorithms, (2) the demonstration of the effectiveness of MADS in achieving faster convergence and lower error rates compared to GA, and (3) the provision of a data-driven, scalable solution for risk prioritization and decision-making in complex and dynamic systems. The results show a significant decrease in convergence error from 0.126 to 0.005, allowing for more precise risk analysis and mitigation strategies.}, archive = {J_EAAI}, author = {Elham Fallah Baghemoortini and Davood Shishebori and Mustafa Jahangoshai Rezaee and Armin Jabbarzadeh}, doi = {10.1016/j.engappai.2025.112104}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112104}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing fuzzy cognitive map convergence through supervised and unsupervised learning algorithms: A case study of operational risk assessment in power distribution networks}, volume = {161}, year = {2025}, }
• (2025). Semi-supervised generative adversarial network for plant leaf disease detection. EAAI, 161, 112103. (www)

  @article{ , abstract = {Fully supervised plant leaf disease segmentation methods based on convolutional neural networks (CNN) require large quantities of labeled images for training, which are time-consuming and labor-intensive to obtain in practical scenarios. Moreover, most diseases of plant leaf exhibit indistinct edge information and background noise interference, which hinders precise disease localization. To overcome these issues, a semi-supervised generative adversarial network (SSGAN) is proposed for plant leaf disease inspection. Firstly, to learn comprehensive semantic features from limited data, a semi-supervised method based on generative adversarial networks is developed to enhance the interaction of features among labeled and unlabeled disease images. Secondly, a boundary feature attention module (BFAM) is proposed to enhance edge detail representation, which makes the model pay more attention to the boundary feature of plant leaf diseases. Finally, a background noise suppression module (BNSM) is proposed to bolster the differentiation between the normal and disease areas so as to alleviate the adverse impact of background noise. The effectiveness of SSGAN is verified on two plant leaf disease datasets. The mean intersection over union (mIoU) on the apple and tomato leaf datasets improves by 2.07 % and 2.28 % respectively, compared with the suboptimal method. The testing results of experiments show that SSGAN achieves great performance on plant leaf disease segmentation under limited labeled data.}, archive = {J_EAAI}, author = {Lixiang Zhao and Jun Hao and Demin Li and Jianbo Yu}, doi = {10.1016/j.engappai.2025.112103}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112103}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Semi-supervised generative adversarial network for plant leaf disease detection}, volume = {161}, year = {2025}, }
• (2025). Industrial device-aided data collection for real-time rail defect detection via a lightweight network. EAAI, 161, 112102. (www)

  @article{ , abstract = {Rail defect detection is challenging due to the diverse and irregular nature of defects, along with the limited availability of high-quality datasets. Existing methods struggle with effectively capturing multi-scale features for proper feature allocation and preserving crucial details in deep networks, leading to incomplete defect representation and reduced accuracy. To address these limitations, we propose Rail Defect Detection Network (REDNet), a lightweight deep learning model specifically designed for real-time rail defect detection in manufacturing and maintenance applications. We design the Multi-Scale Deep Feature Aggregation (MSDFA) module to enhance semantic consistency modeling and achieve more precise feature fusion. We develop the Adaptive Task Decomposition Head (ATDH) to address dynamic feature allocation, and we introduce the Reversible Column Network (RevCol) as the backbone to enhance feature extraction and ensure information reconstruction. Additionally, we developed a high-quality dataset using specialized equipment to address data scarcity and utilized it for training. REDNet achieved a high mean Average Precision at an Intersection over Union (IoU) threshold of 0.50 (mAP50) of 94.1% with exceptional real-time performance at 204.1 frames per second (FPS), while keeping an efficient design of 5.70 million parameters and surpassing state-of-the-art (SOTA) methods in accuracy and speed. These features make it suitable for defect detection, facilitating engineering deployment, and improve quality control in rail manufacturing and maintenance. Generalization tests on the public Microsoft Common Objects in Context (MS COCO) dataset yielded a mean Average Precision across IoU thresholds from 0.50 to 0.95 (mAP50–95) of 46.8%, further confirming the effectiveness of REDNet.}, archive = {J_EAAI}, author = {Qing Dong and Tianxin Han and Gang Wu and Lina Sun and Min Huang and Fu Zhang}, doi = {10.1016/j.engappai.2025.112102}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112102}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Industrial device-aided data collection for real-time rail defect detection via a lightweight network}, volume = {161}, year = {2025}, }
• (2025). A quantitative multi-dimensional resilience framework in process industries. EAAI, 161, 112100. (www)

  @article{ , abstract = {Process industries face significant safety challenges arising from the handling of hazardous materials, tightly coupled operations, and complex system interdependencies. Addressing these challenges requires management strategies that are both proactive and holistic rather than purely reactive. Resilience engineering embodies this approach by emphasizing a system's ability to anticipate disturbances, adapt under stress, and recover effectively. This study introduces a comprehensive, multi-dimensional framework for assessing resilience in process industries. The framework was developed through content analysis of expert interviews, a systematic literature review, and application of the spherical fuzzy Delphi method. Results revealed five primary dimensions of resilience: the organizational dimension, which encompasses leadership commitment, safety culture, and communication processes; the human resources dimension, focusing on workforce competence, training, and well-being; the individual dimension, addressing cognitive readiness, situational awareness, and decision-making skills; the risk management system dimension, which includes hazard identification, risk assessment, emergency preparedness, and continuous monitoring; and the technical dimension, covering equipment reliability, system redundancy, automation safeguards, and maintenance practices. Within these dimensions, factors such as leadership, effective communication, workforce capability, robust risk governance, and technical robustness emerged as particularly influential. By integrating these dimensions into a unified framework, our study advances theoretical understanding of resilience engineering and offers practical guidance for enhancing safety performance in process plants. Furthermore, this framework lays a foundation for future research aimed at developing standardized assessment tools, evaluating long-term outcomes of resilience interventions, and exploring the integration of emerging technologies—such as artificial intelligence, the internet of things, and advanced automation—into resilience engineering practices.}, archive = {J_EAAI}, author = {Mojtaba Emkani and Moslem Alimohammadlou and Esmaeil Zarei and Mehdi Jahangiri and Mojtaba Kamalinia}, doi = {10.1016/j.engappai.2025.112100}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112100}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A quantitative multi-dimensional resilience framework in process industries}, volume = {161}, year = {2025}, }
• (2025). Mutual transfer learning for cuff-less blood pressure estimation using photoplethysmography-based visibility graphs. EAAI, 161, 112099. (www)

  @article{ , abstract = {Cuff-less blood pressure (BP) estimation is critical to cardiovascular disease prevention and management. Photoplethysmography (PPG)-based monitoring technology offers advantages over cuff-based devices, including portability, lower power consumption, and faster measurements. However, current deep learning methods for BP estimation from PPG signals are limited by their analysis only from one-dimensional perspectives, failing to exploration of the underlying physiological patterns of cross-domain visual representations based on higher dimensional perspectives. Furthermore, conventional knowledge distillation techniques necessitate unidirectional knowledge transfer from pre-trained models, rendering it challenging to obtain feedback on the learning state of small networks for optimizing and adjusting the training process. This is inadequate for acquiring a profound understanding of the intricate mapping relationship between PPG and BP values. Therefore, this work presents a novel Transformer-based mutual transfer learning framework (MTL) that estimates BP values from phase-space reconstructed PPG signals using a multi-field complementary approach. The proposed MTL method leverages four phase-space reconstruction techniques to convert PPG signals into visibility graphs (VGs) that provide rich time-variant information. Furthermore, the joint optimization strategy with multiple losses of the soft label and structural knowledge learning enables us to transfer pre-trained knowledge from heterogeneous Transformer models and obtain cumulative multi-field complementary VG features during the fine-tuning process. We evaluate our MTL on three datasets of 1375 subjects using a subject-wise data-splitting paradigm based on five-fold cross-validation, achieving a state-of-the-art performance with estimation errors of 0.50 ± 4.94 millimeter of mercury (mmHg) and 0.21 ± 2.63 mmHg for systolic and diastolic BP, respectively. Our proposed end-to-end MTL offers a computationally efficient solution and elegant generalization ability for BP estimation using PPG-based VGs, providing a novel and innovative approach to BP monitoring that can advance cardiovascular disease prevention and management.}, archive = {J_EAAI}, author = {Chenbin Ma and Zhenchang Liu and Peng Zhang and Lishuang Guo and Haonan Zhang and Zeyu Liu and Guanglei Zhang}, doi = {10.1016/j.engappai.2025.112099}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112099}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Mutual transfer learning for cuff-less blood pressure estimation using photoplethysmography-based visibility graphs}, volume = {161}, year = {2025}, }
• (2025). Groundwater parameters estimation: A hybrid approach of convolutional layers with asynchronous and distributed bio-inspired algorithms. EAAI, 161, 112098. (www)

  @article{ , abstract = {This research focuses on aquifer hydraulic parameters estimation using bio-inspired algorithms since they can tackle groundwater model non-linearities. We propose two novel hybrid frameworks that combine the advantages of convolutional layers (CL) to enhance pattern recognition with heuristic search of Particle Swarm Optimization (PSO) and Differential Evolution (DE) algorithms. These integrations are implemented using an asynchronous and distributed approach to address efficiency issues in large-scale problems, resulting in ADPSO-CL (Asynchronous and Distributed Particle Swarm Optimization with Convolutional Layers) and ADDE-CL (Asynchronous and Distributed Differential Evolution with Convolutional Layers). The distributed method employs virtual machines, where a server generates and assigns particles to workers, which run in parallel with asynchronous iterative solution exchanges. We assess different algorithm configurations in an integrated water management model by coupling two software: Water Evaluation and Planning (WEAP) and MODFLOW. Results indicate that ADPSO-CL outperforms ADDE-CL by demonstrating more stable asynchronous communication, with fewer incomplete experiments (more than one worker was disconnected before completing all iterations), 33% in contrast to 71%. Additionally, produces results closer to the expected values, with mean absolute percentage error (MAPE) values of 78.25% for hydraulic conductivity ( K ) and 55.56% for specific yield ( S y ), compared to 299% and 209% in ADDE-CL. Moreover, ADPSO-CL has the fastest convergence rate, achieving efficient results in about half of the total iterations. This study introduces a novel and scalable architecture for intricate simulation–optimization problems, demonstrating its potential for future applications in real-world water resources planning and management.}, archive = {J_EAAI}, author = {Kiara Tesen and Hermilo Cortés and Sebastián Vicuña and Edmundo Molina-Perez and Francisco Suárez}, doi = {10.1016/j.engappai.2025.112098}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112098}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Groundwater parameters estimation: A hybrid approach of convolutional layers with asynchronous and distributed bio-inspired algorithms}, volume = {161}, year = {2025}, }
• (2025). Forecasting power generation: A novel two-dimensional logistic fractional grey model. EAAI, 161, 112097. (www)

  @article{ , abstract = {Accurate prediction of power generation is essential for the stable operation of power systems, sustained economic growth, and harmonious social development. However, power systems exhibit chaotic characteristics such as complex and nonlinear dynamics during operation; for this reason, this study exploits the fact that two-dimensional logistic mapping helps with the recognition and control of the complex high-dimensional dynamics of the system. Additionally, the fractional-order chaotic system offers a more universal means to describe the chaotic phenomena, etc., and introduces the modelling mechanism of the two-dimensional logistic model and the fractional-order derivative theory into the grey model. A novel two-dimensional logistic fractional-order grey model is developed. The new model can simultaneously analyze two parallel major factors that evolve and are interrelated, and the fractional-order derivatives and nonlinear terms accurately portray the historical state and dynamic evolution process of the system, which can explore the evolution law in the time series of power generation more adequately. Furthermore, the particle swarm optimization algorithm is used to optimize the fractional order derivatives and parameters so that the adjustable model parameters can better capture the change pattern of the original power generation system data. The new model is applied to China's power generation prediction, and its validity is verified from three different perspectives through two different power generation modes as the research objects. Moreover, the new model outperforms the other seven models in its calculations. Finally, the new model effectively forecasts China's power generation for the next five years.}, archive = {J_EAAI}, author = {Mingyue Weng and Huiming Duan and Derong Xie}, doi = {10.1016/j.engappai.2025.112097}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112097}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Forecasting power generation: A novel two-dimensional logistic fractional grey model}, volume = {161}, year = {2025}, }
• (2025). A deep reinforcement learning optimization algorithm based on heterogeneous graph neural network for hybrid flow shop scheduling problem with finite transportation resources. EAAI, 161, 112096. (www)

  @article{ , abstract = {In many actual process industries, transportation equipment represented by automated guided vehicles (AGVs) is widely used. However, current related research rarely considers the coordinated scheduling optimization of AGVs and production equipment. To this end, the hybrid flow shop scheduling problem (HFSP) with finite transportation resources (HFSP-FTR) in process industries is considered in this paper. Since it is necessary to consider the coordinated scheduling optimization of machines and AGVs at the same time, an end-to-end deep reinforcement learning scheduling method based on a heterogeneous graph neural network is proposed. First, a heterogeneous graph model capable of expressing any HFSP-FTR instance is constructed to visually represent the allocation relationship among jobs, machines, and AGVs. Second, a Markov decision process is formulated for solving HFSP-FTR, which specifically encompasses state features based on the heterogeneous graph model and a novel representation method for composite scheduling actions. Then, a novel heterogeneous graph neural network framework that can parallelly embed nodes of various types is proposed, aiming to derive graph-level embeddings for HFSP-FTR instances. Finally, a policy network is designed to obtain the probability distribution of each composite scheduling action being executed, and a deep reinforcement learning framework based on the multi-actor network is proposed for training. The results of numerical simulation experiments conducted on test instances with varying types demonstrate that the proposed method can acquire effective and highly generalized scheduling strategies in terms of makespan performance, outperforming other scheduling algorithms that are widely applied regarding the quality of scheduling solutions and other aspects.}, archive = {J_EAAI}, author = {Yejian Zhao and Xiaochuan Luo and Weixiang Xu and Yulin Zhang}, doi = {10.1016/j.engappai.2025.112096}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112096}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A deep reinforcement learning optimization algorithm based on heterogeneous graph neural network for hybrid flow shop scheduling problem with finite transportation resources}, volume = {161}, year = {2025}, }
• (2025). Unknown fault incremental learning based on shapelet prototypical network for streaming industrial signals. EAAI, 161, 112094. (www)

  @article{ , abstract = {Unknown faults represent faults that have never occurred before, they are constantly emerging due to the changing environments and operations in industrial processes. It is a challenge for existing fault diagnosis methods to continually detect unknown faults and effectively classify known faults in streaming industrial signals. This article proposes an unknown fault incremental learning method for streaming industrial signals. In this work, shapelet prototypical embedding combined with a memory distance matrix is employed to embed streaming industrial signals into a discriminative feature space. Therefore, the category information in the signals can be extracted and is not limited by the size of the sliding window. Besides, a new training paradigm based on meta-learning by sampling simulated-incremental tasks is proposed to obtain generalizable shapelets. Moreover, based on the new training paradigm, the meta-discovery module is proposed to continually detect unknown faults, and the meta-calibrate module can calibrate all prototypes into a distinguishable space. Experiments on the simulated streaming time series, benchmark Tennessee Eastman process, and real-world aluminum electrolysis process illustrate the superiority of the proposed method in terms of accuracy and interpretability. The code is available in https://github.com/XiaoxueWan/UFIL.git .}, archive = {J_EAAI}, author = {Xiaoxue Wan and Lihui Cen and Xiaofang Chen and Yongfang Xie and Zhaohui Zeng}, doi = {10.1016/j.engappai.2025.112094}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112094}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Unknown fault incremental learning based on shapelet prototypical network for streaming industrial signals}, volume = {161}, year = {2025}, }
• (2025). Adaptive attention graph convolution network with normalized embedded gaussian for rapid serial visualization presentation decoding. EAAI, 161, 112093. (www)

  @article{ , abstract = {Graph convolutional networks (GCNs) have been widely used in Brain Computer Interface (BCI) and have shown great prowess in identifying electroencephalogram (EEG) spatiotemporal features. However, most GCNs learn channels topological relationship by fixed adjacency matrix. This lacks connectivity strength information and ignores the data dependency. This paper proposes a data-driven adjacency matrix based on normalized embedded Gaussian function, and constructs a Gaussian-Adaptive Attention Graph Convolution Network (Gaussian-AAGCN). Brain regions connectivity is calculated by normalized embedded Gaussian function, and the topological relationship is adaptively learned by input data in a data-driven manner. This data-driven adaptive adjacency matrix avoids brain activity information loss caused by fixed adjacency matrix and improves the flexibility of graph construction. Convolutional block attention module (CBAM) is introduced to adaptive feature refinement in two independent dimensions, improving model representation ability. Experimental results show that the average area under curve (AUC), true positive rate (TPR) and false positive rate (FPR) of Gaussian-AAGCN on 14 subjects are 93.52 %, 91.59 %, and 4.58 % respectively. Compared to Transformer, Event-Related Potential Capsule Network (ERP-CapsNet), Electroencephalogram Convolutional Neural Network (EEGNet), and Multi-Granularity Information Fusion Network (MGIFNet), the AUC of Gaussian-AAGCN is higher by 18.02 %, 5.32 %, 2.62 %, and 1.82 %, respectively. Using the adaptive adjacency matrix, the model AUC and TPR are increased by about 4.8 % and 7.8 % respectively. After integrating CBAM, the AUC and TPR increased by about 3.5 % and 8 % respectively.}, archive = {J_EAAI}, author = {Mengyuan Zhao and Qingsong Ai and Kun Chen and Quan Liu and Sheng Quan Xie and Li Ma}, doi = {10.1016/j.engappai.2025.112093}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112093}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive attention graph convolution network with normalized embedded gaussian for rapid serial visualization presentation decoding}, volume = {161}, year = {2025}, }
• (2025). GREEN: Graph reasoning enhanced encoder network for social intention-aware forecast of vessel navigating trajectory. EAAI, 161, 112092. (www)

  @article{ , abstract = {Recent studies have demonstrated that graph neural network (GNN) excel in interactive modeling, particularly by enhancing reasoning capabilities through the construction of interaction graphs, and the application of graph convolutional network (GCN). However, these methods remain heavily dependent on prior features, which limits the ability of vessels to accurately infer each other’s decision-making intentions in encounter situations. Therefore, this paper proposes graph reasoning enhanced encoder network (GREEN), which enables social community insight on multi-ships interactions, and latent intention aware for interactive trajectory forecast. The architecture of GREEN includes two layers of trajectory-graph-embedded encoders layer (TGEEL) and multi-graphs-gated fusion layer (MGGFL). In TGEEL, we respectively design intention trajectory generator unit for graph-embedded multi-ships trajectory representation via random walk process, and latent intention generator unit to forecast future navigating intentions via variational autoencoder. In MGGFL, we also design two units, where one unit is in charge of social aware construction and spatio-temporal features extraction based on interactive multi-graphs, and the other one is responsible for gated fusion of multi-scale features and forecast future trajectory using spatio-temporal attention. Experimental results across five datasets were collected from real-world maritime environments, which demonstrated that GREEN achieved improvements of 39.98%, 38.57%, 38.65%, and 63.33% in average displacement error , final displacement error, maximum displacement error, and miss rate, compared with state-of-the-art methods. The paper highlights GREEN robust potential in complex navigational scenarios, and provides pivotal support for advancing the development of intelligent maritime systems.}, archive = {J_EAAI}, author = {Junhao Jiang and Yi Zuo}, doi = {10.1016/j.engappai.2025.112092}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112092}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {GREEN: Graph reasoning enhanced encoder network for social intention-aware forecast of vessel navigating trajectory}, volume = {161}, year = {2025}, }
• (2025). DiffAT: Effective data augmentation with diffusion models for time series forecasting. EAAI, 161, 112091. (www)

  @article{ , abstract = {Data augmentation offers a promising solution to data scarcity in deep learning-based time series forecasting. However, current approaches face dual limitations (1) Hand-designed methods (e.g., cropping/masking): often disrupt the continuity of vital temporal patterns (such as seasonal and trend) by introducing abrupt pattern discontinuities; (2) Generative models often face difficulties in preserving task-critical features that are essential for prediction, especially when aiming to generate diverse augmented series. To tackle these dilemmas, we propose a novel conditional diffusion-based data augmentation framework, named DiffAT, for time series forecasting tasks. DiffAT synergizes: (1) Patch-wise masking reconstruction to capture structural invariants (such as autocorrelation and causality), and (2) encoding hand-designed augmentation prototypes for guiding diversity-preserving generation. DiffAT achieves dual enhancement: maintaining continuity of temporal patterns through progressive denoising process and exposing latent invariant patterns via guided diversity injection. We validate the efficacy of DiffAT through extensive experiments on seven real-world datasets, by comparing DiffAT with six state-of-the-art time series data augmentation methods. The results indicate our method can boost the forecasting performance of Autoformer by up to 6.49 % in 26/28 cases and improve forecasting performance of LightTS by up to 3.11 % in 23/28 cases on 7 real-world benchmarks. Extensive experiments also indicate that DiffAT can improve the accuracy of forecasting models in few-shot scenario (with 1 % training data) in 54/60 cases. We will release the source code upon publication.}, archive = {J_EAAI}, author = {Yang Yu and Ruizhe Ma and Wenbo Gu and Zongmin Ma}, doi = {10.1016/j.engappai.2025.112091}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112091}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {DiffAT: Effective data augmentation with diffusion models for time series forecasting}, volume = {161}, year = {2025}, }
• (2025). Evaluating reinforcement learning-based neural controllers for quadcopter navigation in windy conditions. EAAI, 161, 112090. (www)

  @article{ , abstract = {Accurate quadcopter navigation under windy conditions remains challenging for traditional control methods, especially in the presence of unpredictable wind gusts and strict navigational constraints. This paper evaluates Deep Reinforcement Learning (DRL) based controllers under such conditions, analysing the impact of wind domain randomisation, multi-goal training, enhanced state representations with explicit wind information, and the use of temporal data to capture affecting dynamics over time. Experiments in the AirSim simulator across four trajectories — evaluated under both no-wind and windy conditions — demonstrate that DRL-based controllers outperform classical methods, particularly under stochastic wind disturbances. Moreover, we show that training a DRL agent with domain randomisation improves robustness against wind but reduces efficiency in no-wind scenarios. However, incorporating wind information into the agent’s state space enhances robustness without sacrificing performance in wind-free settings. Furthermore, training with stricter waypoint constraints emerges as the most effective strategy, leading to precise trajectories and improved generalisation to wind disturbances. To further interpret the learned policies, we apply Shapley Additive explanations analysis, revealing how different training configurations influence the agent’s feature importance. These findings underscore the potential of DRL-based neural controllers for resilient autonomous aerial systems, highlighting the importance of structured training strategies, informed state representations, and explainability for real-world deployment.}, archive = {J_EAAI}, author = {Alain Andres and Aritz D. Martinez and Sümer Tunçay and Ignacio Carlucho}, doi = {10.1016/j.engappai.2025.112090}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112090}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Evaluating reinforcement learning-based neural controllers for quadcopter navigation in windy conditions}, volume = {161}, year = {2025}, }
• (2025). Pixel-level semantics boosted fine-grained bird image classification. EAAI, 161, 112089. (www)

  @article{ , abstract = {Fine-grained bird image classification (FBIC) is crucial for endangered bird conservation and biodiversity research. However, existing methods often struggle to capture detailed features and manage the interference caused by complex backgrounds. To address these challenges, we propose a novel Pixel-Level Semantic Boosted Fine-Grained Bird Image Classification (PFIC) framework, which enhances fine-grained bird image classification by incorporating pixel-level semantic information. PFIC consists of two core components: the Grouped Detail Enhancement (GDE) module and the Background–Foreground Enhancement (BFE) strategy. GDE integrates multi-level pixel-level semantic information, derived from a segmentation feature extractor, into classification features via two submodules: grouped aggregation and detail enhancement. This approach enhances the model’s ability to capture fine-grained details. BFE augments training samples by restricting background ranges and applying random shifts to foreground objects, thereby improving the model’s capability to recognize foreground objects in complex environments. Experimental results demonstrate that our method achieves state-of-the-art performance on the CUB-200-2011 and NABirds datasets. Additionally, further experiments on the Stanford Cars dataset validate the framework’s potential for generalization to other fine-grained image classification tasks.}, archive = {J_EAAI}, author = {Haoxiang Ma and Yongjian Deng and Bochen Xie and Jian Liu and Hai Liu and Youfu Li and Zhen Yang}, doi = {10.1016/j.engappai.2025.112089}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112089}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Pixel-level semantics boosted fine-grained bird image classification}, volume = {161}, year = {2025}, }
• (2025). Transformer-based soft sensor of five day biochemical oxygen demand indicators in the municipal wastewater treatment process. EAAI, 161, 112088. (www)

  @article{ , abstract = {Five day biochemical oxygen demand (BOD 5 ) is usually used to measure whether wastewater treatment meets the standard. Previously, the development of machine learning technology was applied to predict BOD 5 for future planning of wastewater treatment. In real wastewater treatment, the monitoring of BOD 5 often lags behind for a long time, so neural network-based soft sensor methods are widely used in this field. However, the currently used BOD 5 soft sensor model has the problem of ignoring the synergistic effect of the model and stable information transmission. In order to monitor the BOD 5 concentration more accurately, a Transformer network model with embedded Long Short-Term Memory (LSTM) network is proposed in this article : LSTM is embedded into the calculation of the transformer multi-head attention, and the output method of the multi-head attention is improved so that the attention heads are output one by one in a serial manner, so it is named LSTM-Serial-Transformer (LSformer). Compared with previous models, LSformer optimizes input information through LSTM and Transformer to achieve synergistic effect, thus better capturing the characteristics of BOD 5 time series data, while reducing the risk of gradient vanishing or gradient exploding during data transmission to improve the stability of the model. Finally, the model was verified to have better soft sensor performance on a benchmark simulation model and a real-world dataset.}, archive = {J_EAAI}, author = {Xiaoling Zhang and Zhi Qi and Peng Chang and Zhiqi Hu}, doi = {10.1016/j.engappai.2025.112088}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112088}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Transformer-based soft sensor of five day biochemical oxygen demand indicators in the municipal wastewater treatment process}, volume = {161}, year = {2025}, }
• (2025). ForgetMe: Benchmarking the selective forgetting capabilities of generative models. EAAI, 161, 112087. (www)

  @article{ , abstract = {The widespread adoption of diffusion models in image generation has increased the demand for privacy-compliant unlearning. However, due to the high-dimensional nature and complex feature representations of diffusion models, achieving selective unlearning remains challenging, as existing methods struggle to remove sensitive information while preserving the consistency of non-sensitive regions. To address this, we propose an Automatic Dataset Creation Framework based on prompt-based layered editing and training-free local feature removal, constructing the ForgetMe dataset and introducing the Entangled evaluation metric. The Entangled metric quantifies unlearning effectiveness by assessing the similarity and consistency between the target and background regions and supports both paired ( Entangled-D ) and unpaired ( Entangled-S ) image data, enabling unsupervised evaluation. The ForgetMe dataset encompasses a diverse set of real and synthetic scenarios, including CUB-200-2011 (Birds), Stanford-Dogs, ImageNet, and a synthetic cat dataset. We apply LoRA fine-tuning on Stable Diffusion to achieve selective unlearning on this dataset and validate the effectiveness of both the ForgetMe dataset and the Entangled metric, establishing them as benchmarks for selective unlearning. Our work provides a scalable and adaptable solution for advancing privacy-preserving generative AI. Code is available at: https://github.com/YuZhenyuLindy/ForgetMe .}, archive = {J_EAAI}, author = {Zhenyu Yu and Mohd Yamani Idna Idris and Pei Wang and Yuelong Xia and Yong Xiang}, doi = {10.1016/j.engappai.2025.112087}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112087}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {ForgetMe: Benchmarking the selective forgetting capabilities of generative models}, volume = {161}, year = {2025}, }
• (2025). A novel data generation scheme for surrogate modelling with deep operator networks. EAAI, 161, 112086. (www)

  @article{ , abstract = {Operator-based neural network architectures such as deep operator networks have emerged as a promising tool for the surrogate modelling of physical systems. In general, the training data for operator surrogate modelling are generated by solving partial differential equations using the finite element method. The computationally intensive nature of data generation is one of the major bottlenecks in deploying these surrogate models, hindering the deployment of these surrogate models in practical applications. In this study, we propose a novel methodology to alleviate the computational burden associated with training data generation for deep operator networks. Unlike the existing literature, the proposed framework for data generation does not use any partial differential equation integration strategy, such as the finite element method. In the proposed strategy, the primary field consistent with the boundary conditions is first generated randomly using Gaussian process regression. Thereafter, from the primary field, the input source field is calculated using second-order accurate finite difference techniques. The computation of the derivatives from the finite difference scheme is significantly less computationally intensive and easier to implement compared to integrating the underlying governing equations, thereby reducing the computational cost associated with generating training datasets. To validate the proposed approach, we employ heat equations as a model problem and develop the surrogate model for numerous boundary value problems.}, archive = {J_EAAI}, author = {Shivam Choubey and Shailendra Rahi and Birupaksha Pal and Manish Agrawal}, doi = {10.1016/j.engappai.2025.112086}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112086}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel data generation scheme for surrogate modelling with deep operator networks}, volume = {161}, year = {2025}, }
• (2025). Interpretable generalization diagnosis: Variational causal disentanglement-based coalitional game attribution network. EAAI, 161, 112084. (www)

  @article{ , abstract = {Interpretability and controllability are the keys to intelligent fault diagnosis in industrial scenarios. However, most intelligent methods have focused on post-hoc feature importance explanations, which cannot constrain the uncontrollable learning behavior or form an interpretable diagnosis process consistent to match expert experience. To address this problem, this paper proposes a variational causal disentanglement-based coalitional game attribution network (VCN) to achieve a controllable and interpretable modeling paradigm that can draw diagnostic conclusions with conforming to expert experience. This model enables independent causal factor disentanglement and model regularization during the training process, as well as fast explanation computation at test time. Variational causal disentanglement is defined and do-operations and variational inference processes are introduced to generate training data with independent causal factors for subsequent attribution. The coalitional game attribution establishes a coalitional game relationship between the model and the independent causal factors, and regularizes the model so that it converges to a state that conforms to logical axioms. Both laboratory and real-world data demonstrate the proposed method enables the provision of interpretable diagnostic conclusions while maintaining controllable performance. The code is available at: https://github.com/cyber-dogy/VCN-code .}, archive = {J_EAAI}, author = {Junwei Gu and Changhua Hu and Yu Wang and Mingquan Zhang and Yanzhuo Lin and Shangjing Peng and Dongdong Li}, doi = {10.1016/j.engappai.2025.112084}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112084}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Interpretable generalization diagnosis: Variational causal disentanglement-based coalitional game attribution network}, volume = {161}, year = {2025}, }
• (2025). A novel photovoltaic cell defect detection method: A deep learning model based on multi-scale enhanced feature extraction. EAAI, 161, 112083. (www)

  @article{ , abstract = {With the increasingly widespread application of clean energy technology, the global demand for photovoltaic cells is gradually increasing, and the status and defects of photovoltaic cells are also being taken seriously. However, there are various types of defects in photovoltaic cells, including those that are difficult to detect, making detection work somewhat challenging. This article proposes a novel improved model for defect detection in photovoltaic cells based on You Only Look Once (YOLO). Firstly, the Multi-Head Self-Attention (MHSA) mechanism is adopted to compensate for the shortcomings of the Cross Stage Partial Bottleneck with 2 Convolutions (C2f) module in channel feature extraction. Secondly, the Receptive-Field Coordinate Attention Convolutional (RFCAConv) operation module is utilized to expand the receptive field of defect detection in the model and increase the range of feature extraction. Finally, the Bi-directional Feature Pyramid Network (BiFPN) and Concat modules are integrated to enhance the feature relationships between the output connections of each module and strengthen the output results. The final experimental results demonstrate that the improved model achieves an accuracy of 90.3%, representing a 5.2% increase compared to the original model. Through experimental verification, the improved model proposed in this paper can not only detect defects in photovoltaic cells in electroluminescence (EL) images but also in infrared thermal images. The model has strong generalization ability.}, archive = {J_EAAI}, author = {Yu Gao and Zhanying Li and Yinghao Zhang and Kangye Zhang and Haoyang Yu}, doi = {10.1016/j.engappai.2025.112083}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112083}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel photovoltaic cell defect detection method: A deep learning model based on multi-scale enhanced feature extraction}, volume = {161}, year = {2025}, }
• (2025). Wind power prediction based on hybrid deep learning and monte carlo simulation. EAAI, 161, 112082. (www)

  @article{ , abstract = {This study proposes a hybrid deep learning model combining Variational Mode Decomposition (VMD), Convolutional Neural Networks (CNN), and a Three-Dimensional Gated Neural Network (TGNN) to enhance wind power prediction accuracy. VMD decomposes wind power time series into intrinsic mode functions, CNN extracts deep features, and TGNN captures temporal dependencies with error feedback control. A Monte Carlo simulation based on the Central Limit Theorem is introduced to evaluate predictive uncertainty and interval coverage. Experimental results from three wind farms in China demonstrate that the proposed model significantly outperforms baseline hybrid models. Specifically, it achieves a coverage probability (CP) of 0.887, an average interval width (AIW) of 200.870, and a normalized mean square error (NMSE) of 0.057. Compared with the VMD-CNN-LSTM model, the proposed VMD-CNN-TGNN reduces root mean error (RMSE) by approximately 50%, indicating its superior accuracy, robustness, and practical value in wind power forecasting.}, archive = {J_EAAI}, author = {Zhiyong Guo and Qiaoli Han and Fangzheng Wei and Wenkai Qi}, doi = {10.1016/j.engappai.2025.112082}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112082}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Wind power prediction based on hybrid deep learning and monte carlo simulation}, volume = {161}, year = {2025}, }
• (2025). A parallel memetic algorithm for qubit mapping on noisy intermediate-scale quantum machines. EAAI, 161, 112081. (www)

  @article{ , abstract = {Quantum computer technologies are advancing rapidly, but current devices are still limited to the realm of Noisy Intermediate Scale Quantum (NISQ) systems. Due to the limited set of gates and scarce physical qubit connectivity, most quantum circuit-based programs need to be transpiled to be executed on such hardware. This transpilation, which consists in converting the quantum circuit into a hardware-compliant one includes the critical qubit mapping step, where logical qubits are mapped onto physical qubits to meet the hardware connectivity constraints. In this paper, we investigate the application of Artificial Intelligence (AI) to the qubit mapping problem. More exactly, we propose a Parallel Memetic Algorithm for Qubit Mapping (PMA-QM), designed to tackle this challenge efficiently. Using a fine-tuned parallel model to accelerate this inherently computationally expensive hybrid approach, our algorithm takes problem-specific knowledge into account to optimize circuit depth, reducing execution time, and minimizing error rates consequently. PMA-QM has been experimented using various medium-to-large scale circuit benchmarks. PMA-QM outperforms the widely used SWAP-based BidiREctional (SABRE) algorithm, delivering consistently better solutions.}, archive = {J_EAAI}, author = {Jérôme Rouzé and Nouredine Melab and Jan Gmys and Daniel Tuyttens}, doi = {10.1016/j.engappai.2025.112081}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112081}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A parallel memetic algorithm for qubit mapping on noisy intermediate-scale quantum machines}, volume = {161}, year = {2025}, }
• (2025). Enhancing non-orthogonal multiple access systems: A reconfigurable machine learning classification approach. EAAI, 161, 112080. (www)

  @article{ , abstract = {Non-Orthogonal Multiple Access (NOMA) is being considered as a Multiple Access (MA) technique for the next-generation systems, due to factors such as a high per-user spectral resource allocation, grant-free transmission, support for millimeter Waves (mmWaves) and massive-MIMO (mMIMO). A practical limitation of the NOMA systems is imperfect Successive Interference Cancellation (SIC); and the involved decoding delay. This work develops a data-driven Machine Learning (ML) model providing the functionality of a SIC receiver. While most of the Deep Learning (DL) algorithms have high complexity and training times, the given ML approach utilizes the received symbol as the only primary predictor. The other two predictors are derived from the primary predictor, and utilized for the Power-Domain (PD)-NOMA symbol-decoding process. The model utilizes a developed low-complexity NOMA-ML-based Decoder (MLbD) dataset for the same. The extensive test simulations confirm the reconfigurable ML-based receiver to be at par with the existing Maximum Likelihood (MLH) decoder in terms of decoding accuracy. Still, the former supports its integration into the next-generation systems due to its reconfigurable nature and removes the drawbacks related to SIC in the NOMA systems.}, archive = {J_EAAI}, author = {Saurabh Srivastava and Rampravesh Kumar and Prajna Parimita Dash}, doi = {10.1016/j.engappai.2025.112080}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112080}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing non-orthogonal multiple access systems: A reconfigurable machine learning classification approach}, volume = {161}, year = {2025}, }
• (2025). Collaborative energy-saving path planning of unmanned surface vehicle cluster based on multi-head attention mechanism and multi-agent deep reinforcement learning. EAAI, 161, 112078. (www)

  @article{ , abstract = {This paper proposes a Multi-Agent Deep Reinforcement Learning (MADRL) algorithm to address the overlooked issue of energy consumption optimization in existing cluster path planning methods, enabling collaborative energy-efficient path planning for Unmanned Surface Vehicle (USV) clusters. Meanwhile, it provides an alternative MADRL approach to address practical engineering challenges. We not only consider the problems of avoiding danger, reaching the target point and avoiding path conflicts between USV, but also further consider the energy consumption and speed planning of USV cluster. Specifically, firstly, we establish the USV motion model and the USV cluster conflict model, and propose a new energy consumption model, which considers the relationship among speed, marine environment and propulsion load. Secondly, we propose a multi-head local state relevance capture mechanism-multi-agent proximal policy optimization (MLSRC-MAPPO) algorithm. This algorithm can use multi-head attention mechanism (MHA) to capture the potential dependencies between USV, thus enhancing the convergence performance of multi-agent. Finally, in order to reduce the training difficulty, we propose a multi-dimensional action space method for action networks. The experimental results demonstrate that the proposed multi-dimensional action space method has achieved significant success in the lightweighting of the action network: the number of network parameters was reduced by 89.10%, and the computational complexity was decreased by 88.94%, significantly enhancing training efficiency. Meanwhile, the MLSRC-MAPPO algorithm, by incorporating a multi-head attention mechanism, greatly improved the convergence performance of the multi-agent system. In test scenarios with both identical and different starting points, the method reduced energy consumption by 26.24% and 38.47% respectively, fully validating its effectiveness and superiority. Furthermore, comparative experiments with existing cluster energy-saving path planning methods show that the proposed method exhibits clear advantages in terms of energy consumption and path planning efficiency, further verifying its superiority. The corresponding code for this paper is as follows: https://github.com/xhpxiaohaipeng/Multi_USV_Trajectory_energy_plan}, archive = {J_EAAI}, author = {Haipeng Xiao and Lijun Fu and Chengya Shang and Yunfeng Lin and Longfei Yue and Yaxiang Fan}, doi = {10.1016/j.engappai.2025.112078}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112078}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Collaborative energy-saving path planning of unmanned surface vehicle cluster based on multi-head attention mechanism and multi-agent deep reinforcement learning}, volume = {161}, year = {2025}, }
• (2025). Identification of surface subsidence risk in deep foundation pits using a mamba fusion model. EAAI, 161, 112077. (www)

  @article{ , abstract = {This study proposes a novel artificial intelligence–driven neural network model, CNN-Mamba-LSTM-SA, for predicting surface settlement induced by deep foundation pit excavation. The model integrates Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), a Self-Attention (SA) mechanism, and the Mamba architecture to capture both spatial and long-range temporal dependencies in multi-source monitoring data. Bayesian optimization is employed for hyperparameter tuning, and Variational Mode Decomposition (VMD) is used for data denoising, resulting in improved prediction accuracy. To enhance model interpretability, Shapley Additive Explanations (SHAP) are applied to identify key deformation drivers, revealing groundwater level and building settlement as the most influential factors. Model performance is validated using monitoring data from the Nanjing Gemini excavation project, where it achieves superior results compared to conventional models. The CNN-Mamba-LSTM-SA model reduces Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE) by up to 69.77 %, 62.45 %, and 79.45 %, respectively. Further analysis through ablation experiments confirms the contribution of each module. Interestingly, the combined removal of CNN and SA results in greater performance degradation than the sum of their individual effects. Finally, a risk warning framework is developed to enable the dynamic transformation of predicted and observed settlement values into actionable risk levels. This integrated artificial intelligence approach offers a robust and interpretable tool for managing excavation-induced risks in urban geotechnical engineering.}, archive = {J_EAAI}, author = {Chenhe Ge and Pengfei Li and Mingju Zhang and Meng Yang}, doi = {10.1016/j.engappai.2025.112077}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112077}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Identification of surface subsidence risk in deep foundation pits using a mamba fusion model}, volume = {161}, year = {2025}, }
• (2025). Risk management in maintenance processes: A spherical fuzzy-based failure mode and effect analysis approach in the glass processing industry. EAAI, 161, 112076. (www)

  @article{ , abstract = {Quality management systems are essential tools that aim to increase the competitiveness of businesses and ensure customer satisfaction by creating reliable quality control processes. In this study, the importance of the "risk-based thinking" approach of the International Organization for Standardization (ISO) 9001:2015 standard within the scope of quality management systems was emphasized, and a risk assessment was made for the maintenance process of a business in the glass processing sector. In order to eliminate the deficiencies of the failure mode and effect analysis (FMEA) method, the integration of spherical fuzzy sets (SFSs) and the Technique of Order Preference Similarity to the Ideal Solution (TOPSIS) method was used. Within the scope of the study, 17 different risks were determined by five experts from the maintenance, repair, and quality assurance departments. The experts evaluated these identified risks according to occurrence, severity, and detection factors. Then, these risks were ranked using the Spherical Fuzzy TOPSIS (SF-TOPSIS) method. Finally, different scenarios were created, and their results were discussed to provide a more comprehensive and sensitive risk management approach. This study focuses on maintenance processes in the glass processing industry as a case study. However, the risks related to the maintenance process defined in the study (e.g., machine failures, maintenance inefficiencies, spare parts shortages) are common in the manufacturing sector. Therefore, it can be applied to the sector where the application was carried out and to many different sectors and enterprises.This methodology also serves as a guide for businesses that want to manage process risks within the scope of ISO 9001:2015.}, archive = {J_EAAI}, author = {Buse Duygu Dağıdır and Barış Özkan}, doi = {10.1016/j.engappai.2025.112076}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112076}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Risk management in maintenance processes: A spherical fuzzy-based failure mode and effect analysis approach in the glass processing industry}, volume = {161}, year = {2025}, }
• (2025). Few-shot augmentation based on variational auto-generative adversarial network with moving losses: Application to the variable stiffness prediction in composites. EAAI, 161, 112075. (www)

  @article{ , abstract = {The dataset of carbon fiber reinforced plastics (CFRP) materials presents characteristics of high-dimensional, low-rank, and sparse, which pose difficulties in the combination of mechanical modeling. In this paper, a Variational Auto-Generative Adversarial Network (VAGAN) with moving losses is proposed as a data augmentation method, which extends the size of the CFRP dataset covering components, processes, elasticity, and strengths factors, and increases the information conveyed in the surrogate modeling. A compression strength prediction model for CFRP laminates was constructed by combining the component, process, and mechanical tensor with a neural network optimized by the search algorithm. Combined with the data augmentation strategy, not only was the amount of data expanded, but the prediction accuracy was also significantly improved. The allowable value of compression strength is analyzed and calculated by the predicted values, which brings direct benefits in simplifying the test.}, archive = {J_EAAI}, author = {Zhicen Song and Yunwen Feng and Cheng Lu and Jiaqi Liu}, doi = {10.1016/j.engappai.2025.112075}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112075}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Few-shot augmentation based on variational auto-generative adversarial network with moving losses: Application to the variable stiffness prediction in composites}, volume = {161}, year = {2025}, }
• (2025). Multivariate time series forecasting based on temporal decomposition and graph neural network. EAAI, 161, 112074. (www)

  @article{ , abstract = {It is quite challenging to forecast the Multivariate Time Series (MTS) accurately due to the high dimensionality of MTS and the entangled correlation between variables. Recently, graph-based networks have been demonstrated to be an effective model to handle the complex correlations between MTS. However, all existing graph-based methods construct the graph model of the MTS using only the shallow correlations from the raw MTS data, ignoring the deep-rooted correlations hidden in the features. In this paper, we propose for the first time to construct a comprehensive graph model of MTS that incorporates both shallow correlations from raw data and hidden correlations from decomposed temporal properties. Then, we propose a novel graph-based MTS forecasting framework, which optimizes the graph structure jointly with the model parameters. By doing so, the graph structure can adaptively model the correlations of MTS at a deep level, while the joint optimization can make the constructed graph compatible with the forecasting tasks of MTS, contributing to a globally optimal solution. Finally, we conduct extensive experiments on seven real-world datasets, the results demonstrate the superiority of our method on MTS forecasting over the state-of-the-art baselines. The source codes of the experiments with datasets are available at https://github.com/ironweng/MF-TDGNN .}, archive = {J_EAAI}, author = {Yan Qiao and Pei Zhao and Junjie Wang and Rongyao Hu and Minyue Li and Xinyu Yuan and Meng Li and Zhenchun Wei and Cuiying Feng}, doi = {10.1016/j.engappai.2025.112074}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112074}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multivariate time series forecasting based on temporal decomposition and graph neural network}, volume = {161}, year = {2025}, }
• (2025). A unified framework for lung disease screening using chest X-ray images. EAAI, 161, 112073. (www)

  @article{ , abstract = {Chest X-ray (CXR) imaging is a crucial diagnostic tool for identifying respiratory diseases. The existing methods face challenges such as misclassification caused by overlapping radiologic patterns and the shortage of trained radiologists. However, previous studies have utilized deep learning (DL) models to overcome these limitations, but they struggle with imbalanced datasets, and noise sensitivity. In addition, most of the studies are based on the binary classification of CXR images. Thus, the proposed framework integrates segmentation and multi-class classification for improved generalization and robustness. In this work, artificial intelligence techniques like encoder–decoder and ensemble learning are utilized for segmentation and classification of CXR images that offers reliable, and efficient solution. The output of the presented framework classifies the CXR images into four different classes, along with their corresponding class score. The proposed classification model is ensemble of three pre-trained DL architectures. The last five layers of these base models are fine tuned to match the classification process and combined through a multi-layer perceptron classifier for improved accuracy. The proposed model achieves overall accuracy of 88.98%, and area under curve value of 0.9753, outperforming several state-of-the-art models. It is a lightweight model and demonstrates significant robustness under noisy conditions compared to other models. Moreover, the proposed segmentation and ensemble models are trained and tested on different datasets to obtain greater robustness. The data augmentation is also employed to address class imbalance nature of dataset and enhance the generalization capability. Further, statistical analysis is performed to present the comprehensive comparison among the models.}, archive = {J_EAAI}, author = {Bhavana Singh and Pushpendra Kumar}, doi = {10.1016/j.engappai.2025.112073}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112073}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A unified framework for lung disease screening using chest X-ray images}, volume = {161}, year = {2025}, }
• (2025). A hybrid bayesian model updating and non-dominated sorting genetic algorithm framework for intelligent mix design of steel fiber reinforced concrete. EAAI, 161, 112071. (www)

  @article{ , abstract = {Steel fiber reinforced concrete (SFRC) improves the strength and toughness of conventional concrete, but the high cost and carbon footprint of fibers challenge the balance among performance, cost and sustainability. To address this, an intelligent mix design framework is proposed to optimize compressive and splitting tensile strengths, cost and emissions. Based on 671 experimental records, posterior models were built using Markov Chain Monte Carlo sampling and Bayesian model updating, enabling accurate strength predictions. Compared to traditional regression methods, R 2 scores improved by 15.7 % and 12.4 %, confirming its predictive advantage. Cost-wise, materials dominate, while emissions mainly arise from production, transport and mixing. A non-dominated sorting genetic algorithm identified optimal designs under given constraints. Results show that reducing water-to-cement and aggregate-to-cement ratios, and increasing sand ratio and fiber reinforcement index, enhances SFRC strength. Larger coarse aggregates reduce compressive strength but have limited effect on tensile strength. Optimization suggests potential cost and emission reductions of up to 60 %. Moreover, for compression-prone components, fiber use is inefficient due to high cost and emissions, whereas for crack-resistant or strength-balanced elements, fiber inclusion offers a more sustainable alternative to merely lowering the water-to-cement ratio. The proposed framework enables tailored SFRC mix designs, guiding the efficient use of steel fibers.}, archive = {J_EAAI}, author = {Yong Yu and Jie Su and Bo Wu}, doi = {10.1016/j.engappai.2025.112071}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112071}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A hybrid bayesian model updating and non-dominated sorting genetic algorithm framework for intelligent mix design of steel fiber reinforced concrete}, volume = {161}, year = {2025}, }
• (2025). Three-dimensional human pose estimation based on multi-scale spatial–temporal transformer. EAAI, 161, 112068. (www)

  @article{ , abstract = {Recently, transformer-based methods have become dominant in the domain of three-dimensional (3D) human pose estimation, yet the U-net model based on convolutional neural networks (CNN) struggles to model long temporal sequences, and sequence-to-frame (Seq2frame) and sequence-to-sequence (Seq2seq) approaches often fail to preserve dependencies at the start and end of sequences. To address these challenges, this paper proposes a Multi-Scale Spatial–Temporal Transformer network (MSST). This network utilizes Sequence Padding Module (SPM) to extract edge features of the first and last frames, and employs Spatial–Temporal Transformer (STT) to model the spatial–temporal correlations of keypoints. Additionally, we design a Multi-Scale Module (MSM) that analyzes the human skeletal topology to extract multi-scale features of keypoints, local information, and global information, and fuse semantic information at different scales. Finally, we utilize regression heads to project the processed keypoint feature information into 3D space. We conduct quantitative evaluations on two benchmark datasets using four evaluation metrics and design multiple sets of comparative experiments to validate the effectiveness of the proposed modules. Experimental results demonstrate that the proposed network achieves excellent performance.}, archive = {J_EAAI}, author = {Xiaogang Song and Yongxin Cui and Jichen Chen and Xinhong Hei}, doi = {10.1016/j.engappai.2025.112068}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112068}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Three-dimensional human pose estimation based on multi-scale spatial–temporal transformer}, volume = {161}, year = {2025}, }
• (2025). SharpEdge: High-quality data-driven monocular depth estimation for enhanced boundary precision. EAAI, 161, 112067. (www)

  @article{ , abstract = {While existing monocular depth estimation methods have achieved commendable performance, they often fall short in accurately distinguishing object boundaries. This deficiency largely stems from the inherent noise in dataset acquisition, such as unclear edges and missing depth information. To address these challenges, this paper introduces a novel, high-quality, data-driven monocular depth estimation method tailored for autonomous driving. The approach significantly enhances depth predictions with clearer object boundaries and reduced noise, making it well-suited for real-time, safety-critical applications. Central to our approach is the Self-Adaptive Consistency Filtering mechanism, which dynamically selects high-quality training samples, ensuring that the model learns from the most reliable data and reducing the impact of noise. Additionally, we introduce a Dual-Prior Learning strategy that combines geometric and semantic edge priors. Unlike traditional methods that rely solely on raw depth maps, our approach enhances boundary detection by providing detailed guidance on object contours. This leads to more accurate depth estimation, especially in complex regions where other methods struggle. Empirical evaluations on popular benchmark datasets show that our approach leads to performance improvements of 1.2% on the autonomous driving dataset KITTI and 1.8% on the indoor scene dataset NYU. Compared with recent state-of-the-art methods such as DPT and NewCRFs, our approach achieves superior performance, particularly in recovering fine object boundaries and maintaining spatial consistency across diverse scenes. These results highlight the strong generalization ability of our method, demonstrating that it can enhance depth estimation quality across diverse environments — improving edge precision and spatial coherence, which are critical for autonomous vehicles navigating both complex and dynamic scenarios.}, archive = {J_EAAI}, author = {Mengke Song and Luming Li and Xu Yu and Chenglizhao Chen and Shanchen Pang}, doi = {10.1016/j.engappai.2025.112067}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112067}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {SharpEdge: High-quality data-driven monocular depth estimation for enhanced boundary precision}, volume = {161}, year = {2025}, }
• (2025). A systematic review on vision-based gaze estimation: Advance in computer vision and deep learning. EAAI, 161, 112066. (www)

  @article{ , abstract = {Eye gaze estimation is all about figuring out where a person is looking by analyzing eye movements. This is usually done using cameras or eye-tracking devices, a key technique in computer vision. Gaze estimation has a wide range of uses—from helping people interact with computers more naturally (HCI), monitoring drivers, enhancing virtual reality experiences, and reading emotional cues in affective computing. Deep learning has improved how accurately and reliably these systems work in recent years. Models like Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Generative Adversarial Networks (GANs) have made it easier to detect and interpret eye movements in complex settings. This review closely examines how gaze estimation works, the datasets commonly used, and both traditional and deep learning-based methods. It also points out what's working well, where current methods fall short, and what researchers still need to solve. The goal is to offer helpful insights for anyone working to improve or apply gaze tracking technologies in real-world scenarios.}, archive = {J_EAAI}, author = {Sapna Singh Kshatri and Deepak Singh}, doi = {10.1016/j.engappai.2025.112066}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112066}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A systematic review on vision-based gaze estimation: Advance in computer vision and deep learning}, volume = {161}, year = {2025}, }
• (2025). A novel framework for improving railway driver performance based on emotional intelligence and job-driven factors: An artificial neural network method. EAAI, 161, 112065. (www)

  @article{ , abstract = {Efficient transportation systems critically rely on human factors to ensure safety, reliability, and service quality. This study introduces a novel framework for evaluating and improving the performance of the Tehran Urban and Suburban Railway Company (TRS) drivers by focusing on the role of emotional intelligence (EI) and job-driven (JD) factors. The core of the proposed method is a two-stage hybrid approach: first, an Artificial Neural Network (ANN) is used to model driver performance using EI and JD factors as inputs and outputs, respectively; then, Data Envelopment Analysis (DEA) validates the results and benchmarks performance levels. Data were collected from 146 TRS drivers through a standardized questionnaire. Sensitivity analysis and statistical tests evaluated the impact of EI and JD factors on driver performance. Additionally, a SWOT (Strengths, Weaknesses, Opportunities, Threats) analysis translated the findings into practical recommendations for system improvement. The findings reveal high job satisfaction and low job stress levels among TRS drivers, highlighting the significant contribution of EI and JD factors. By combining data-driven prediction, performance benchmarking, and strategic analysis, this study introduces a novel, human-centered framework for driver Performance Evaluation (PE)—one that fills methodological gaps in the literature and supports enhancing safety, efficiency, and passenger satisfaction in railway systems.}, archive = {J_EAAI}, author = {Narges Hajloo and Behnaz Salimi and Mahdi Hamid and Masoud Rabbani}, doi = {10.1016/j.engappai.2025.112065}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112065}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A novel framework for improving railway driver performance based on emotional intelligence and job-driven factors: An artificial neural network method}, volume = {161}, year = {2025}, }
• (2025). An efficient fusion-based deep learning framework for land use and land cover image clustering. EAAI, 161, 112061. (www)

  @article{ , abstract = {Land use and land cover (LULC) analysis is vital for understanding spatial dynamics and informing environmental management, urban planning, and sustainable development. Traditional approaches, such as manual surveys and conventional image clustering methods, often face limitations in scalability and adaptability. This paper presents a novel deep learning framework that combines the Vision Transformer (ViT) and Variational Autoencoder (VAE) to extract complementary feature representations for LULC image clustering. The ViT tokenizes image patches to capture high-level semantic features, while the VAE models latent structures to integrate contextual and structural information. To further improve clustering performance, the framework incorporates Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction followed by k -means++ clustering, enabling a scalable and robust solution for diverse datasets. Experiments on multiple datasets, including the Urban Atlas LULC 2018 dataset and recent LULC maps of Japan and Vietnam, demonstrate the framework’s superior ability to capture complex LULC patterns compared to traditional methods. The datasets and source code will be made publicly available at https://github.com/ClarkDinh/LULCMiner . This framework has broad applications across geospatial and remote sensing engineering, civil and environmental engineering, agricultural planning, transportation, and urban development.}, archive = {J_EAAI}, author = {Tai Dinh and Dat Tran and Zdena Dobešová and Huynh Van Hong and Daniil Lisik and Rameesh Khan}, doi = {10.1016/j.engappai.2025.112061}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112061}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An efficient fusion-based deep learning framework for land use and land cover image clustering}, volume = {161}, year = {2025}, }
• (2025). Optimizing multi-objectives urban siting of hydrogen refueling dispenser using fuzzy NSGA-II: A case study in fez, morocco. EAAI, 161, 112060. (www)

  @article{ , abstract = {To support the growing need for sustainable urban transport, it is essential to reduce the environmental impact of vehicles; this study focuses on optimizing hydromobility through the strategic integration of hydrogen dispensers (HDs) into existing locations, rather than deploying standalone hydrogen refueling stations (HRSs). This approach minimizes costs while enhancing accessibility, addressing the current limitations in hydrogen vehicle (HV) adoption. A novel method combining the non-dominated sorting genetic algorithm II (NSGA-II) with fuzzy logic and Pareto front analysis is proposed to identify optimal HD locations using a real case study of Fez city in Morocco. The results provide actionable strategies for implementing HDs in a way that balances efficiency, accessibility, and budget constraints. The study demonstrates how intelligent planning can support the transition to cleaner energy solutions in urban mobility.}, archive = {J_EAAI}, author = {Soukayna Abibou and Dounia El Bourakadi and Ali Yahyaouy and Hamid Gualous}, doi = {10.1016/j.engappai.2025.112060}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112060}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Optimizing multi-objectives urban siting of hydrogen refueling dispenser using fuzzy NSGA-II: A case study in fez, morocco}, volume = {161}, year = {2025}, }
• (2025). Synthetic data enhancement using diffusion models for improved miscanthus identification and bioenergy extraction. EAAI, 161, 112059. (www)

  @article{ , abstract = {Miscanthus, a high-yielding perennial grass pivotal for bioenergy production, requires precise species identification to optimize bioenergy extraction. However, limited annotated spectral datasets hinder robust classification model development. To address this challenge, the paper proposes a novel diffusion probabilistic model tailored for near-infrared spectral synthesis. Unlike conventional generative approaches, the proposed model integrates a bidirectional gated recurrent unit-based temporal encoder and one dimension convolutional neural networks within a diffusion framework, augmented by a spectral attention module to prioritize critical absorption bands. This architecture uniquely addresses the sequential dependencies and subtle biochemical variations inherent in near-infrared spectral, enabling high-fidelity generation of diverse synthetic data. The diffusion process is optimized through a hybrid loss function combining variational lower bound training with mean squared error for pixel-level fidelity and maximum mean discrepancy for distributional alignment. Evaluated on 517 near-infrared spectral samples across three Miscanthus species, the proposed model outperforms traditional variational autoencoders , generative adversarial networks, and standard diffusion models in terms of sample authenticity and diversity. Incorporating synthetic data enhanced the accuracy, precision, and recall of downstream classifiers by 10%–15%, with the convolutional neural networks attaining 87% accuracy using hybrid real-synthetic training data. Remarkably, even with 50% synthetic data substitution, classification accuracy remained robust at 75%, demonstrating the model’s efficacy in mitigating data scarcity and advancing precision agriculture for bioenergy optimization.}, archive = {J_EAAI}, author = {Xinyue Wang and Xiangdong Chen and Jun Jiang and Ronggao Gong and Biao Wang}, doi = {10.1016/j.engappai.2025.112059}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112059}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Synthetic data enhancement using diffusion models for improved miscanthus identification and bioenergy extraction}, volume = {161}, year = {2025}, }
• (2025). Multi-order feature learning with global–local attention for real-time X-ray security inspection. EAAI, 161, 112058. (www)

  @article{ , abstract = {X-ray security inspection in high-throughput environments such as airports faces fundamental challenges stemming from the intrinsic complexity of baggage scanning. These challenges include feature entanglement from overlapping objects, visual distortions due to fixed imaging angles, inconsistent pseudo-color representations caused by varying X-ray absorption, and the difficulty of detecting prohibited items across multiple scales. To tackle these challenges, we introduce Multi-Order Gated Network (MoGNet), a lightweight transformer-based architecture that integrates three core innovations. First, a Multi-order Gated Aggregation Block designed for efficient multi-order feature extraction. Second, a Global–Local Self-Attention mechanism that enhances differentiation between foreground and background elements. Finally, a DynamicFusion module for adaptive integration of multi-scale features. Comprehensive evaluations on five challenging datasets establish our proposed framework, MoGNet, as the new state-of-the-art (SOTA). The model demonstrates superior performance, achieving mean Average Precision (mAP) scores at a 50% Intersection over Union (IoU) threshold (mAP 50 ) of 75.4%, 91.9%, 91.2%, 96.6%, and 80.0%, respectively. This high accuracy is maintained while operating at an efficient 64.1 Frames Per Second (FPS). These comprehensive experimental results demonstrate its remarkable capability to optimize the balance between computational efficiency and detection accuracy, establishing it as a preferable solution for real-time contraband detection in practical security screening.}, archive = {J_EAAI}, author = {Ling Guo and Yangbin Xu and Shouhong Chen}, doi = {10.1016/j.engappai.2025.112058}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112058}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-order feature learning with global–local attention for real-time X-ray security inspection}, volume = {161}, year = {2025}, }
• (2025). A stress-dependent strength neural network model for predicting the true triaxial strength of rocks. EAAI, 161, 112057. (www)

  @article{ , abstract = {Understanding the true triaxial strength of rocks is essential for safe underground engineering, yet existing empirical and data-driven models often fail to capture the nonlinear effects of the intermediate principal stress σ 2 . This study proposes a stress-dependent strength neural network (SDSNN) that integrates physically informed constraints, including monotonicity and boundary conditions, as well as an exponential adaptive weighting strategy to balance data and constraint losses. Cohesion and internal friction angle are used as input features, replacing conventional reliance on uniaxial strength. Compared with a purely data-driven neural network (DDNN) and a constraint-aware variant without adaptive weighting (SDSNN#0), SDSNN achieves significantly better predictive accuracy and robustness across test sets, representative rock types, and five-fold cross-validation. In particular, it maintains consistent strength trends and improved stability across six representative rocks with diverse mechanical properties. Importantly, the model maintains high performance even when trained only on low-to-mid σ 2 data and tested on high σ 2 conditions—demonstrating strong generalization under extrapolation. This capability has received relatively limited attention in data-driven models and is particularly valuable in practical scenarios where high- σ 2 test data are limited or difficult to obtain. Furthermore, ablation analysis demonstrates that removing physical constraints leads to a notable decrease in model accuracy, underscoring the importance of incorporating strength variation characteristics into the model. SDSNN shows particular advantage under boundary stress conditions and when facing noisy or sparse datasets, indicating its potential to serve as a robust and interpretable tool for true triaxial strength prediction in geotechnical applications.}, archive = {J_EAAI}, author = {Tianzhi Yao and Yunpeng Gao and Jianhai Zhang and Ru Zhang and Li Qian and Qijun Hu and Xianliang Wang and Feng Jiang}, doi = {10.1016/j.engappai.2025.112057}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112057}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A stress-dependent strength neural network model for predicting the true triaxial strength of rocks}, volume = {161}, year = {2025}, }
• (2025). Multi-source contrastive cluster center method for cross-domain bearing fault identification. EAAI, 161, 112056. (www)

  @article{ , abstract = {Because of the complexity and time-varying attributes of the exterior surroundings, rolling bearings commonly operate under variable working conditions at any time. Therefore, there is a multi-source domain adaptation problem that involves multiple source domains and a target domain. In this circumstance, identifying faults directly in a multi-source domain through a model built in a single source domain will also lead to limited model generalization ability, i.e., due to the incompleteness of the training sample size caused by the singularity of the domain quantity limited access to fault knowledge in a single source domain, the established model may be prone to over-fitting, thus whose generalization ability has been decreased under complex and variable working conditions to a certain extent. Hence, this paper has proposed a Multi-source Contrastive Cluster Center (MS3C) Method for addressing the aforementioned issues. Experimental findings on two datasets have suggested that MS3C has not only considered the domain shifts of the same classes between different source domains and the target domain but also adaptively aligned the feature distributions of the same classses in different source domains, therefore, MS3C has a higher identification rate, a better clustering and classification performance and a superior convergence.}, archive = {J_EAAI}, author = {Pengfei Chen and Lizhen Wu and Rongzhen Zhao and Kongyuan Wei and Yuqiao Zheng and Linfeng Deng and Yongfei Zhang and Mingkuan Shi and Zhuo Chen}, doi = {10.1016/j.engappai.2025.112056}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112056}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-source contrastive cluster center method for cross-domain bearing fault identification}, volume = {161}, year = {2025}, }
• (2025). An enhanced encoder–decoder network for temporomandibular joint segmentation in magnetic resonance images. EAAI, 161, 112054. (www)

  @article{ , abstract = {Accurate segmentation of the condyle, articular disc, and articular eminence from magnetic resonance imaging (MRI) is essential for the diagnosis and treatment planning of temporomandibular joint (TMJ) disorders. However, current deep learning methods for TMJ segmentation face significant challenges: (1) anatomical structures of the TMJ exhibit complex inter-patient variability; (2) structural boundaries are often ambiguous; and (3) large inter-slice spacing in MRI makes it difficult to fully utilize spatial context, limiting the effectiveness of both three-dimensional and two-dimensional approaches. To address these challenges, we propose an edge-enhanced TransUNet-based encoder–decoder framework that effectively integrates local edge features and global contextual information to obtain accurate and robust TMJ segmentation. Specifically, a cross-slice attention transformer (CAT) is introduced to capture inter-slice dependencies, addressing the challenge of large slice spacing and enhancing the utilization of contextual information across slices. Moreover, a feature enhancement module (FEM) is designed to explicitly fuse edge information, facilitating accurate localization where boundaries are blurred or indistinct. In addition, an attention gate (AG) mechanism adaptively highlights salient anatomical structures and suppresses irrelevant background, improving the model’s overall focus and resistance to noise. Evaluated on the private TMJ MRI dataset, our method achieves Dice similarity coefficients (DSC) of 0.922, 0.834, and 0.837 for the condyle, articular disc, and articular eminence, respectively, outperforming baseline and comparison methods. Further validation on a public knee MRI dataset demonstrates the good generalizability of the proposed method. These results demonstrate robust and precise TMJ segmentation, supporting reliable automated analysis in clinical settings.}, archive = {J_EAAI}, author = {Yilin Hu and Yunan Zhang and Wei Tang and Jixiang Guo}, doi = {10.1016/j.engappai.2025.112054}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112054}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {An enhanced encoder–decoder network for temporomandibular joint segmentation in magnetic resonance images}, volume = {161}, year = {2025}, }
• (2025). Three-way decision method integrating probabilistic linguistic term sets and prospect theory in large-scale scenarios. EAAI, 161, 112053. (www)

  @article{ , abstract = {Probability linguistic term sets (PLTSs), owing to their ability to effectively represent the uncertainty and complexity of decision information, have been widely adopted in large-scale multiattribute decision-making (MADM) scenarios. However, existing decision-making methods in the PLTS environment typically suffer from information loss in similarity computation, subjectivity in weight determination, and inadequate consideration of decision-makers’ behavioral psychology, which limits their applicability in real-world complex situations. To address these issues, this paper proposes an innovative hybrid three-way decision (TWD) method that integrates prospect theory with multiobjective optimization on the basis of the ratio analysis (MOORA) approach. Compared with existing studies, the main contributions of this paper include (1) an objective weight determination mechanism based on the information entropy of PLTSs, which improves the scientific validity and computational efficiency of weight allocation compared with subjective or complex traditional methods (2) a novel direct similarity degree based on the probability linguistic Jensen–Shannon (PLJS) divergence, which effectively overcomes the information loss caused by indirect distance measures and achieves accurate similarity characterization between PLTSs (3) a new conditional probability determination strategy utilizing θ -level similarity classes that enhances the dynamic adaptability and accuracy of conditional probability estimation and (4) the integration of prospect theory and the MOORA method within the TWD framework, enabling hierarchical classification of alternatives while fully reflecting decision-makers’ psychological and behavioral characteristics, thereby facilitating the classification and ranking of alternatives. Furthermore, through comparative analysis with large-scale real-world air quality assessment data and mainstream PLTS-based decision-making methods, the effectiveness and rationality of the proposed method are comprehensively verified, demonstrating its superiority, robustness, and practical application value.}, archive = {J_EAAI}, author = {Zhaxi Pahua and Haidong Zhang and Yizhu Cairang and Yanping He}, doi = {10.1016/j.engappai.2025.112053}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112053}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Three-way decision method integrating probabilistic linguistic term sets and prospect theory in large-scale scenarios}, volume = {161}, year = {2025}, }
• (2025). Adaptive feature-level fusion of manifold and deep learning for robust multi-view face recognition. EAAI, 161, 112052. (www)

  @article{ , abstract = {Robust face recognition across varying poses, lighting conditions, and viewpoints remains a significant challenge in engineering applications such as surveillance, authentication, and human–computer interaction. This paper presents Dynamic Manifold–Deep Learning Fusion (DM-DLF), an artificial intelligence (AI)-based framework that integrates manifold learning with deep neural networks using an adaptive feature-level weighting strategy. By combining global semantic features and local geometric structures, the proposed AI model enhances identity recognition under real-world multi-view scenarios. Experimental results show that DM-DLF surpasses recent transformer-based and graph-based models, achieving up to 94.00 % classification accuracy. The method also reduces prediction error and improves training efficiency. These findings confirm DM-DLF as a powerful AI solution for multi-view face recognition in engineering systems where labeled data is limited and view diversity is high.}, archive = {J_EAAI}, author = {Faraein Aeini}, doi = {10.1016/j.engappai.2025.112052}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112052}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Adaptive feature-level fusion of manifold and deep learning for robust multi-view face recognition}, volume = {161}, year = {2025}, }
• (2025). Enhancing long-term load forecasting with convolutional informer-based hybrid model. EAAI, 161, 112051. (www)

  @article{ , abstract = {Long-term load forecasting (LTLF) is essential for energy management but challenged by the complexity of non-stationary time series. The Informer model struggles to capture localized peak–valley patterns, while Variational Mode Decomposition (VMD) faces issues with feature complexity. This study proposes a hybrid framework integrating VMD, Informer, and a Convolutional Long Short-Term Memory (CNN-LSTM) module for accurate LTLF. VMD decomposes non-stationary load data into multi-scale intrinsic mode functions, refined through spectral and autocorrelation analyses to ensure robust feature extraction. The Informer employs sparse self-attention for efficient long-sequence modeling, with CNN-LSTM enhancing the decoder to capture localized temporal dynamics. Experiments on non-stationary load time series across multiple prediction horizons demonstrate that the proposed framework significantly improves forecasting accuracy and robustness compared to baseline models, including Informer and its derivatives. By excelling in complex load pattern prediction, the framework supports efficient grid scheduling and resource optimization in energy systems.}, archive = {J_EAAI}, author = {Bin Sun and Xudong Chen and Tao Shen and Liyao Ma}, doi = {10.1016/j.engappai.2025.112051}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112051}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing long-term load forecasting with convolutional informer-based hybrid model}, volume = {161}, year = {2025}, }
• (2025). Fault diagnosis of high-voltage three-phase asynchronous motors using residual neural networks and bidirectional gated recurrent units. EAAI, 161, 112049. (www)

  @article{ , abstract = {High-voltage three-phase asynchronous motors are extensively used in industrial production; however, rotor bar breakage faults are characterized by strong concealment and severe consequences, making early and accurate diagnosis difficult with traditional methods. To address this issue, this paper proposes a deep learning model based on a Residual Neural Network (ResNet) and a Bidirectional Gated Recurrent Unit (Bi-GRU) for motor fault diagnosis. The proposed model integrates the spatial feature extraction advantages of ResNet with the temporal modeling capabilities of Bi-GRU, simultaneously capturing spatiotemporal information from voltage and current signals. The model was trained and tested on a large-scale dataset comprising 324,000 data points under nine distinct operating conditions. Experimental results demonstrate that the proposed method achieves over 99.83 % in key metrics such as accuracy, precision, recall, and F1-score, significantly outperforming traditional approaches and other comparative models. Additionally, it maintains stable performance under varying operating conditions, demonstrating strong robustness and generalization capability. Further ablation experiments also validated the effectiveness of each module within the proposed model. This study indicates that the application of deep learning in industrial motor fault diagnosis has promising prospects and practical value.}, archive = {J_EAAI}, author = {Huihui Yang and Yuxin Wu}, doi = {10.1016/j.engappai.2025.112049}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112049}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Fault diagnosis of high-voltage three-phase asynchronous motors using residual neural networks and bidirectional gated recurrent units}, volume = {161}, year = {2025}, }
• (2025). Zero-shot image translation via query compensation and style enhancement. EAAI, 161, 112048. (www)

  @article{ , abstract = {In zero-shot image translation methods based on diffusion models, two prevalent challenges are the loss of identity information and insufficient stylization. To address these issues, we propose a zero-shot image translation method. Specifically, we use the edit-friendly noise space for image inversion. By this way, the identity of the input image is largely preserved. Additionally, we introduce an identity compensation mechanism by injecting the source query vectors into the denoising process. Furthermore, to tackle the model’s insufficient stylization ability, we propose cross-attention style modulation (CSM) to transfer style information from the reference to the input image. Finally, to further enhance the style effect of the translated image, we design initial latent self-enhancement (ILS), style supplementation (SS), and style alignment (SA) strategies. Our zero-shot image translation method does not require training samples, optimization, or fine-tuning. We achieve this by manipulating the self-attention features of a pre-trained diffusion model in a manner analogous to cross-attention—by replacing the key and value vectors of the input image with those of the reference image during the denoising process. Extensive in-domain and cross-domain translation experiments demonstrate the effectiveness of our method across a wide range of object categories and show strong robustness to variations in object shape, size, posture and instance between the input and reference image. The proposed method achieves competitive performance in identity preservation, as measured by the mean Intersection over Union (mIoU), and attains the best style transfer in terms of Single Image Fréchet Inception Distance (SIFID). In addition, qualitative comparisons demonstrate that our approach significantly outperforms the baselines.}, archive = {J_EAAI}, author = {Heng Zhang and Yi-Jun Yang and Wei Zeng}, doi = {10.1016/j.engappai.2025.112048}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112048}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Zero-shot image translation via query compensation and style enhancement}, volume = {161}, year = {2025}, }
• (2025). Camouflaged object detection with boundary localization in complex backgrounds. EAAI, 161, 112047. (www)

  @article{ , abstract = {The primary challenge of Camouflaged Object Detection (COD) lies in the high similarity between the target and the complex background, making it difficult for the human eye to distinguish them. Based on the phenomenon that human attention shifts between the target and the background when observing objects, we propose a network model named MENet. This model adopts a three-stage decoupled architecture of “localization-interaction-fusion.” In the localization stage, we utilize an attention mechanism-based backbone network (Pyramid Vision Transformer V2, abbreviated as PVT-V2) to generate multi-level features, which can initially locate the target area. In the interaction stage, we design a Contour-Aware Edge Module (CAEM) and an Area Decoder (AD) to capture the target edges and background information, respectively, thereby achieving precise localization of the target boundary and reducing interference from background noise. Furthermore, we developed a Boundary Guidance Module (BGM) that effectively injects boundary cues and relevant background information separately into the multi-level features, enhancing the model’s ability to detect target edges in complex backgrounds. In the fusion stage, we design two Feature Fusion Modules (FFM and KFFM) to effectively merge multi-level features with precise boundaries and de-noised features, thereby enhancing the prediction performance of camouflaged objects. Extensive experiments on three challenging benchmark datasets demonstrate that our MENet outperforms many existing state-of-the-art methods. Our method leverages artificial intelligence (AI) techniques to improve the accuracy of camouflaged object and pest detection in complex visual environments. Our code is publicly available at: https://github.com/yang19950966666/MENet .}, archive = {J_EAAI}, author = {Guangjian Zhang and Zhengming Yang and Yong Wang and Yuliang Chen and Duoqian Miao}, doi = {10.1016/j.engappai.2025.112047}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112047}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Camouflaged object detection with boundary localization in complex backgrounds}, volume = {161}, year = {2025}, }
• (2025). A grid-based boundary sharpening clustering algorithm. EAAI, 161, 112045. (www)

  @article{ , abstract = {To address the clustering problem for arbitrary shapes, in this paper, we propose a Grid-based Boundary Sharpening clustering algorithm called as “GBSharp”. This method is grounded in morphology and relies on two fundamental morphological operations: dilation and erosion. The main innovations of the proposed algorithm lie in two aspects. Firstly, we further introduce the concepts of inward dilation and bridge erosion based on the basic morphological operations to reduce the impact of the chain effect. Secondly, a unique indexing structure is designed specifically for non-empty cells in high dimensional space. In addition, to tackle the complex conditional judgments encountered in high-dimensional scenarios, we further utilize the inversion method for bridge-erosion operation. Experiments conducted on synthetic datasets and real-world datasets further validate the effectiveness and efficiency of the proposed algorithm.}, archive = {J_EAAI}, author = {Lin Ma and Qijing Yan and Mengxia Lv and Tiefeng Ma and Mingchang Cheng}, doi = {10.1016/j.engappai.2025.112045}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112045}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A grid-based boundary sharpening clustering algorithm}, volume = {161}, year = {2025}, }
• (2025). Advances in physics-informed neural networks for solving complex partial differential equations and their engineering applications: A systematic review. EAAI, 161, 112044. (www)

  @article{ , abstract = {Most physical and engineering problems can be described by partial differential equations (PDEs), which are typically solved using numerical methods such as the finite difference method and the finite element method. However, conventional numerical discretization approaches face significant challenges in terms of computational efficiency and convergence speed when dealing with complex nonlinear PDEs, such as high-dimensional nonlinear PDEs, stiff PDEs, PDEs with complex boundary conditions or irregular geometries, and multi-scale PDEs. Recently, physics-informed neural networks (PINNs) have emerged as a transformative methodology for solving complex PDEs by integrating physical laws intrinsically into deep learning architectures. While PINNs effectively overcome mesh dependency and dimensionality constraints inherent in traditional numerical methods, they still encounter persistent challenges related to training convergence and generalization robustness. This paper aims to present a comprehensive review of the state-of-the-art developments in PINNs for solving complex PDE problems. The core ideas, network architectures, and generic implementation frameworks, along with associated open-source Python libraries, are first introduced in detail. Furthermore, a systematic taxonomy of optimization techniques is provided, covering hyperparameter selection, adaptive sampling strategies, physics-constrained loss formulations, hybrid differentiation approaches, and architectural innovations. Subsequently, various coping strategies and research advancements of PINNs in addressing complex nonlinear PDE problems are thoroughly discussed. Real-world engineering applications are then reviewed across multiple domains, including cosmology and quantum mechanics, materials science and manufacturing, fluid mechanics, energy systems, biological and environmental sciences, and power and information technologies. Finally, this paper discusses the current challenges and limitations of PINNs in solving complex PDEs and outlines potential directions for future research. By addressing the current limitations and pursuing targeted improvements in architectures, training, interpretability and generalization, PINNs can become a powerful tool in engineering and scientific applications.}, archive = {J_EAAI}, author = {Jiangtao Guo and Hao Zhu and Yujie Yang and Chenrui Guo}, doi = {10.1016/j.engappai.2025.112044}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112044}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Advances in physics-informed neural networks for solving complex partial differential equations and their engineering applications: A systematic review}, volume = {161}, year = {2025}, }
• (2025). DPMF-net: A dual-path perceptive multi-stage fusion network for skin lesion segmentation. EAAI, 161, 112043. (www)

  @article{ , abstract = {Accurate segmentation of skin lesions in dermoscopic images is crucial for skin cancer detection and treatment. Despite progress in deep learning-based methods, challenges remain due to diverse skin lesion shapes, colors, and blurred boundaries. We propose a novel Dual-path Perceptive Multi-stage Fusion Network (DPMF-Net) for skin lesion segmentation. DPMF-Net integrates multiple feature refinement modules. It aims to gradually optimize lesion representations by leveraging the dual-path framework to perceive high-level contextual information. The Spatial Frequency Dual-path Cascaded Perception Module (SFDCP) synergizes spatial and frequency domains to model long-range dependencies and suppress noise, enhancing perception of low-contrast lesions. Subsequent to the SFDCP, the Spatial Channel Dual-path Parallel Perception Module (SCDPP) employs entropy-driven attention and multi-granularity convolutions in skip connections to dynamically select informative channels and extract lesion details across spatial scales. To verify the efficacy of our proposed DPMF-Net, extensive experimental assessments are carried out across four challenging datasets. The outcomes of these quantitative and qualitative experiments confirm that our approach significantly outperforms current state-of-the-art methods in terms of all evaluation metrics.}, archive = {J_EAAI}, author = {Yuling Huang and Yaoyao Ma and Jing Wang and Chao Xu and Zhiwei Fan and Di Wu}, doi = {10.1016/j.engappai.2025.112043}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112043}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {DPMF-net: A dual-path perceptive multi-stage fusion network for skin lesion segmentation}, volume = {161}, year = {2025}, }
• (2025). Gradient surgery: A necessity for robust test-time adaptation for detecting casting defects. EAAI, 161, 112039. (www)

  @article{ , abstract = {Casting defects pose a significant challenge in the manufacturing industry, leading to material waste, production inefficiencies, and compromised product quality. While deep learning models have shown promise in automating defect detection, their effectiveness is often constrained by domain shifts and variability in real-world data distributions. In this work, we propose Bayesian Test-Time Adaptation (BTTA), a novel framework designed to enhance the robustness and adaptability of machine learning models in such dynamic environments. Unlike traditional Test-Time Adaptation (TTA) methods, our approach employs gradient-guided diversification with Stein Variational Gradient Descent (SVGD) to explore diverse optimization paths. Experimental results on benchmark datasets, including CIFAR-10-C , Casting Defects , and GDXray , demonstrate significant performance improvements across key metrics. Notably, the framework achieves an average accuracy improvement of 2-3% under severe corruption levels and excels in cross-domain generalization, highlighting its ability to handle diverse and unseen defect categories. This dynamic adaptability not only addresses the limitations of static models but also offers a practical and cost-effective solution for real-time defect detection in industrial settings. Our study underscores the potential of BTTA to transform quality assurance processes, ensuring reliable performance across varying operational conditions without the need for extensive retraining or large annotated datasets. The codebase for BTTA is available on: https://github.com/afsharshamsi/GradSurgery .}, archive = {J_EAAI}, author = {Afshar Shamsi and Rejisa Becirovic and Hamid Alinejad-Rokny and Arash Mohammadi and Ahmadreza Argha}, doi = {10.1016/j.engappai.2025.112039}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112039}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Gradient surgery: A necessity for robust test-time adaptation for detecting casting defects}, volume = {161}, year = {2025}, }
• (2025). Enhancing real-time detection transformer for small floating targets. EAAI, 161, 112038. (www)

  @article{ , abstract = {Addressing the challenge of detecting small floating target in complex water environments, where it is difficult to balance real-time performance and end-to-end capabilities, this study introduces a specialized model called the Enhancing Real-Time Detection Transformer (ERT-DETR). This model integrates a Dynamic Feature Pyramid Network (DFPNet) with a Micro-Attention Module (MAM) to achieve refined feature extraction and enhanced object recognition mechanisms. Additionally, it incorporates an Inner Intersection over Union (Inner-IoU) auxiliary bounding box loss function, which accelerates model convergence and improves bounding box accuracy. Experiments conducted on the Floating Object in Water Image Dataset (FloW-Img) demonstrate that the ERT-DETR model achieves a 92.6% Average Precision (AP) and 118.84 Frames Per Second (FPS), outperforming the baseline Real-Time Detection Transformer (RT-DETR) by 4.3% and 19.8%, respectively. The ERT-DETR model’s precision and real-time performance in dynamic water surface environments surpass those of the most advanced You Only Look Once (YOLO) and Detection Transformer (DETR) series detectors. This model is significant for enhancing small floating target detection capabilities in complex water environments and can be extended to applications in marine management, environmental monitoring, and vessel tracking.}, archive = {J_EAAI}, author = {Guobing Xie and Xinran Wu and Jiefeng Shi and Yixin Su and Binghua Shi}, doi = {10.1016/j.engappai.2025.112038}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112038}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Enhancing real-time detection transformer for small floating targets}, volume = {161}, year = {2025}, }
• (2025). Non-contact weight intelligent estimation based on yak skeleton localization. EAAI, 161, 112036. (www)

  @article{ , abstract = {Weight estimation is a vital method for monitoring the growth and health of yaks, however, traditional techniques-such as relying on herders’ experience or using weighbridge-are labor-intensive, time-consuming and pose safety risks. Currently, many studies have shown that yak body size can be an effective indicator of weight. With the advancement of computer vision, non-contact weight estimation has become increasingly feasible for livestock. Yet, studies focusing on yak weight estimation, particularly in high-altitude plateau regions, remain limited. Herein, we propose a novel weight estimation approach based on deep learning and binocular vision technology to address this gap. The method involves four main steps: (1) yak image acquisition, (2) skeletal key point localization, (3) body size calculation (4) weight estimation using Gaussian process regression. To enhance practicality and mobility, we also developed two edge-intelligent devices: an intelligent inspection vehicle and a handheld detection unit, enabling convenient and non-invasive weight estimation. Our models are trained and tested on a yak dataset collected by our team on the Tibetan Plateau. Experimental results demonstrate the effectiveness of our approach, achieving an Mean Absolute Percentage Error (MAPE) of 0.12 percent, a Mean Absolute Error (MAE) of 25.4 kilograms (kg) and a Coefficient of determination ( R 2 ) value of 0.72. It not only provides a new technical solution for the yak industry but also provides innovative insights for advancing intelligent animal husbandry. The code and data can be accessed at https://github.com/FeiWang-swun/YakWeight .}, archive = {J_EAAI}, author = {Fei Wang and Xinghua Zou and Zhijiang Chen and Qi Tang and Tianshuo Li and Shuiying Wang and Lijun Yang and Dongming Tang}, doi = {10.1016/j.engappai.2025.112036}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112036}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Non-contact weight intelligent estimation based on yak skeleton localization}, volume = {161}, year = {2025}, }
• (2025). Comparison of the performances of artificial intelligence bots using continuous intuitionistic fuzzy evaluation based on distance from average solution method. EAAI, 161, 112033. (www)

  @article{ , abstract = {The rapid evolution of artificial intelligence (AI) has introduced novel opportunities and challenges in various fields. In this study, we present a pioneering approach known as Continuous Intuitionistic Fuzzy (CINFU) Evaluation based on Distance from Average Solution (EDAS), an innovative extension of the EDAS method tailored to Continuous Intuitionistic Fuzzy Sets. This methodology is designed to compare the performance of AI tools. The capabilities of AI bots have been examined through their success rates in various tasks and uncertainty levels in decision-making processes. The study aims to evaluate the effectiveness of different models in decision-making processes by analyzing the performances of AI bots such as Chat Generative Pre-trained Transformer (ChatGPT), Bard, and Claude based on both objective measurements and fuzzy evaluation criteria. The comparison focuses on key performance criteria such as Bots Triggered, User Engagement, Message Click-Through Rate, Chat Handoff, User Retention, Bounce Rate & Dwell Time, Leads Captured, and Customer Satisfaction Score. Ultimately, the validity and robustness of the approach have been tested with sensitivity analysis.}, archive = {J_EAAI}, author = {Nurşah Alkan and Umut Aydın and Akın Menekşe and Cengiz Kahraman}, doi = {10.1016/j.engappai.2025.112033}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112033}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Comparison of the performances of artificial intelligence bots using continuous intuitionistic fuzzy evaluation based on distance from average solution method}, volume = {161}, year = {2025}, }
• (2025). A global linear attention incorporated video transformer for robust sintering condition recognition. EAAI, 161, 112032. (www)

  @article{ , abstract = {Robust and accurate sintering condition recognition is a fundamental yet critical issue in the design of image-based intelligent combustion control systems. However, owing to the weak texture and fast changing characteristics of flame videos, capturing the condition indicator using existing gradient-based methods is challenging. To address this issue, we propose a global linear attention incorporated video transformer model for sintering condition recognition. First, to reduce the prediction error and uncertainty, the spatial-temporal features are extracted to describe the dynamic characteristics of the flame video streams based on the video shifted window (Swin) Transformer architecture. Next, to address the problem that the local attention strategy used in the Video Swin Transformer is insufficient for global flame feature extraction, we propose a Video Linear Attention block that obtains the global attention as a supplement. Extensive experiments conducted on a real-world rotary kiln sintering dataset demonstrate the effectiveness of our approach, achieving an overall accuracy of 97.76% and an F1-score of 95.30%. Compared to the Video Swin Transformer model, these results represent improvements of 2.00% in accuracy and 4.96% in F1-score, respectively. This research is particularly significant in the context of real-time identification of combustion process conditions, optimization of control parameters, and realization of more stable and efficient combustion process control.}, archive = {J_EAAI}, author = {Leyuan Wu and Junlin Wu and Dingxiang Wang and Qiang Fu}, doi = {10.1016/j.engappai.2025.112032}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112032}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A global linear attention incorporated video transformer for robust sintering condition recognition}, volume = {161}, year = {2025}, }
• (2025). Markerless gait analysis for parkinson’s disease diagnosis: A study on machine learning integration and features explainability. EAAI, 161, 112031. (www)

  @article{ , abstract = {Parkinson’s disease is a common neurodegenerative disorder with progressive loss of dopaminergic and other subcortical neurons that significantly impair motor functions, necessitating accurate diagnostic and monitoring techniques. Researchers commonly use gait analysis to estimate gait parameters to classify diseases and their progression. Traditional gait analysis, while effective, typically requires expensive, specialized equipment and lacks the flexibility for widespread clinical use. Furthermore, many existing studies on markeless approaches do not use public datasets, giving comparable and benchmarking results. Additionally, these methods often do not incorporate explainability, leaving a gap in understanding which specific gait features most significantly impact diagnostic outcomes. To address this issue, we propose a novel system to support diagnosis and analysis of the gait of Parkinson’s disease. The two-step approach involves pose extraction with vision-based pose estimation and classification employing machine learning techniques based on the subject’s temporal, kinematic, and symmetric characteristics. The explainability technique used in this study shows the important role of features such as the knee angle, cadence, and double support duration in influencing diagnostic outcomes. The system shows promising results for non-invasive, low-cost Parkinson’s disease diagnostics, reaching an accuracy of 0.95, a sensitivity of 0.90, a specificity of 0.97, a Matthew’s Correlation Coefficient (MCC) of 0.88, an Area Under the Curve (AUC) of 0.99.}, archive = {J_EAAI}, author = {Cesare Davide Pace and Alessandro Marco De Nunzio and Claudio De Stefano and Francesco Fontanella and Mario Molinara}, doi = {10.1016/j.engappai.2025.112031}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112031}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Markerless gait analysis for parkinson’s disease diagnosis: A study on machine learning integration and features explainability}, volume = {161}, year = {2025}, }
• (2025). Offshore wind power multi-step forecasting based on multi-scale attention mechanism fusion network. EAAI, 161, 112026. (www)

  @article{ , abstract = {The nonlinear, unstable, and multi-scale characteristics of offshore wind power present significant challenges for grid scheduling and power forecasting. To address this issue, this paper proposes a multi-step forecasting hybrid model for offshore wind power based on the sequence-to-sequence (Seq2Seq) architecture, which is named Multi-scale, Denoising, Attention Fusion Network (mDAFNet). The model integrates a multi-level learnable wavelet decomposition network, frequency-domain interpolation denoising technique, and dual-attention mechanism, and improves forecasting performance through joint high- and low-frequency dual-branch predictions. The process begins by improving data quality through outlier cleaning and feature selection based on grey correlation analysis and mutual information. Subsequently, a multi-level learnable wavelet decomposition network is introduced to extract high- and low-frequency subsequences from the frequency domain and optimize parameters dynamically through backpropagation. To address the noise impact on high-frequency subsequences, frequency domain interpolation denoising technique is applied for noise reduction. Then, a sequence-to-sequence model based on the dual-attention mechanism is used to forecast the denoised high-frequency subsequences, while a Long Short-Term Memory (LSTM) neural network predicts the low-frequency subsequences. Finally, the dual-branch forecasting network enables joint modeling of high- and low-frequency subsequences. We conducted experiments on two offshore wind power datasets. The results indicate that, compared to baseline models, the proposed mDAFNet reduces Mean Absolute Error, Root Mean Square Error, and Mean Absolute Percentage Error by an average of 33.94 %, 36.62 %, and 39.84 %, respectively, while significantly improving forecasting accuracy and stability. This study provides a new approach and perspective for offshore wind power forecasting.}, archive = {J_EAAI}, author = {Yizhuang Xiong and Wenbo Wang}, doi = {10.1016/j.engappai.2025.112026}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112026}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Offshore wind power multi-step forecasting based on multi-scale attention mechanism fusion network}, volume = {161}, year = {2025}, }
• (2025). Weighted multi-granularity fuzzy probabilistic rough set based on semi-overlapping function and its application in three-way decision. EAAI, 161, 112023. (www)

  @article{ , abstract = {This paper introduces a novel model of fuzzy probabilistic rough sets and develops an interesting three-way decision (TWD) applicable to multi-attribute group decision-making and classification problems. First, we simultaneously consider the cut-levels of fuzzy relations and decision-making fuzzy sets, investigating a new fuzzy probabilistic rough set model. This model employs semi-overlapping functions as the range of the membership functions of the fuzzy set, enhancing interpretability since the semi-overlapping functions can act as the truth tables of fuzzy logic. Second, by assigning different weights to various granular fuzzy relations, we propose a weighted multi-granularity fuzzy probabilistic rough set model. It is proven that the proposed (weighted multi-granularity) fuzzy probabilistic rough set model encompasses many existing models as special cases, thereby providing a very broad framework for rough set theory. Third, by integrating the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method with Bayesian theory, we develop a method for computing the threshold of fuzzy probabilistic rough sets and construct a new TWD. Sorting and classification rules and algorithms for TWD are provided. Finally, we perform parameter analysis and comparative experiments on nine public datasets to verify the effectiveness and rationality of our TWD. The experimental results demonstrate that our approach exhibits superior fault tolerance and noise reduction capabilities. Specifically, the introduction of semi-overlapping functions enhances our TWD model’s performance, with A c c u r a c y improving the most on the Heart dataset, reaching 9.37%, and the F 1 score increasing the most on the Heart and Plrx datasets, reaching 5.3%. In summary, this paper offers a novel solution for applying artificial intelligence in decision-making scenarios.}, archive = {J_EAAI}, author = {Xinru Li and Lingqiang Li and Chengzhao Jia}, doi = {10.1016/j.engappai.2025.112023}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112023}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Weighted multi-granularity fuzzy probabilistic rough set based on semi-overlapping function and its application in three-way decision}, volume = {161}, year = {2025}, }
• (2025). Feature-driven double deep Q-network with iterated greedy for intelligent scheduling optimization in reentrant hybrid flow shops. EAAI, 161, 112012. (www)

  @article{ , abstract = {The reentrant hybrid flow shop scheduling problem (RHFSP) poses significant challenges due to its complex structure and limited research on model accuracy, solution space exploration, and adaptive strategy selection. To address these gaps, this paper proposes a novel feature-driven double deep Q-network with iterated greedy (FD3QNIG) algorithm. First, a five-dimensional mixed-integer linear programming (MILP) model is developed to significantly improve modeling accuracy. Second, a feature-driven initialization strategy (FDNEH) enhances the quality of initial solutions, while an adaptive historical information-driven destruction and reconstruction strategy (AHDDR) effectively balances exploration and exploitation during the search process. Third, multi-scale local search strategies are employed, including the first application of comprehensive exploration-driven mandatory operations local search (CED_MOLS) to RHFSP, substantially deepening the search capability. Fourth, an adaptive strategy selection mechanism based on double deep Q-Network (DDQN_ASS) dynamically guides the algorithm toward more efficient decision-making. Extensive experiments on 285 benchmark instances demonstrate that FD3QNIG achieves a 53 %–94 % improvement in average relative percentage increase ( A R P I ) over state-of-the-art methods, confirming its effectiveness and robustness.}, archive = {J_EAAI}, author = {Chexiang Li and Yuyan Han and Yuting Wang and Yiping Liu and Biao Zhang and Leilei Meng}, doi = {10.1016/j.engappai.2025.112012}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112012}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Feature-driven double deep Q-network with iterated greedy for intelligent scheduling optimization in reentrant hybrid flow shops}, volume = {161}, year = {2025}, }
• (2025). A semi-supervised framework for generating multi-dimensional taxonomies from asset maintenance documents. EAAI, 161, 112010. (www)

  @article{ , abstract = {The operation and maintenance of buildings generate large volumes of unstructured textual data, such as inspection reports and service requests. These records contain valuable insights that can support fault detection, cost tracking, and resource planning. However, existing classification approaches often rely on static, expert-defined labels that fail to reflect the complexity of real-world maintenance operations. This paper introduces a hybrid framework that combines sentence embedding, clustering, topic modeling, and network modularization to uncover recurring patterns in maintenance text. The extracted patterns are then reviewed and refined by facility management experts to develop a multi-dimensional taxonomy model tailored to operational needs. The methodology is applied to a case study involving over 30,000 work orders. The results demonstrate how the proposed system captures fine-grained details such as system type, failure mode, and required trade expertise. A proof-of-concept software tool, developed in collaboration with facility managers, showcases the practical value of the taxonomy in enabling data-driven decision-making, such as identifying cost drivers and recurring issues. Additionally, the resulting taxonomy models serve as effective prompts for zero-shot text classification, enabling large language models to classify new maintenance records without requiring retraining or labeled data. This approach provides a scalable and adaptable foundation for text classification systems in asset management.}, archive = {J_EAAI}, author = {Soroush Sobhkhiz and Tamer El-Diraby}, doi = {10.1016/j.engappai.2025.112010}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112010}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A semi-supervised framework for generating multi-dimensional taxonomies from asset maintenance documents}, volume = {161}, year = {2025}, }
• (2025). A comprehensive review of facial beauty prediction using deep learning techniques. EAAI, 161, 112009. (www)

  @article{ , abstract = {Facial beauty prediction (FBP) is an emerging area of artificial intelligence (AI) that focuses on the development of models that analyze facial features to assess beauty, based on human perception. This task is particularly challenging due to the subjective nature of beauty and limited resources. Deep learning methods have proven their exceptional ability to capture complex features and are therefore well suited for FBP tasks. This paper reviews recent advances in FBP, focusing on deep learning techniques and benchmark datasets. A proposed taxonomy organizes the main methods according to their design and applications, and comparative analyzes highlight trends and the performance of different models. Finally, the study outlines future research directions to drive progress in this evolving field.}, archive = {J_EAAI}, author = {Djamel Eddine Boukhari and Fadi Dornaika and Ali Chemsa and Abdelmalik Taleb-Ahmed}, doi = {10.1016/j.engappai.2025.112009}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112009}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A comprehensive review of facial beauty prediction using deep learning techniques}, volume = {161}, year = {2025}, }
• (2025). Maximum relevant minimum redundant multi-label feature selection using ant colony optimization. EAAI, 161, 112007. (www)

  @article{ , abstract = {Multi-label learning tasks involve instances that may belong to multiple categories simultaneously, making feature selection particularly challenging in high-dimensional feature spaces. Existing multi-label feature selection methods often suffer from limitations such as high computational complexity, inadequate handling of feature redundancy, and insufficient modelling of label dependencies. To overcome these challenges, we propose a novel framework called Maximum Relevant Minimum Redundant Multi-Label Feature Selection (MR2MLFS), which integrates a two-layer graph representation with a modified Ant Colony Optimization (ACO) strategy. The first graph layer clusters correlated features using Louvain community detection, while the second constructs a meta-graph to model inter-cluster relationships. ACO then explores this structure, favouring the selection of highly relevant and non-redundant features. To reduce computational overhead, we introduce an information-theoretic metric that estimates both feature-label relevance and feature-feature redundancy, eliminating the need for repeated classifier training during the search. We evaluated the proposed method on ten benchmark multi-label datasets using several multi-label classifiers. Experimental results show that the proposed method outperforms six state-of-the-art methods across multiple evaluation metrics, achieving an average relative improvement of 5–12 % while reducing feature dimensionality by up to 80 %. These results confirm the method's robustness, efficiency, and effectiveness in multi-label feature selection.}, archive = {J_EAAI}, author = {Mohammad Hatami and Parham Moradi and Sadegh Sulaimany and Mahdi Jalili}, doi = {10.1016/j.engappai.2025.112007}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112007}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Maximum relevant minimum redundant multi-label feature selection using ant colony optimization}, volume = {161}, year = {2025}, }
• (2025). A two-stage model for unified sentence- and document-level biomedical event extraction. EAAI, 161, 112001. (www)

  @article{ , abstract = {Biomedical event extraction, a cornerstone of information extraction, has increasingly attracted attention within the biomedical research community. Moreover, it is a highly complex task, which not only deals with many sub-tasks but also involves nested events. Currently, the research on biomedical event extraction, whether pipelined model or joint method, needs to be processed for each sub-task. The process of processing each sub-task one by one lead to the degradation of event extraction performance. In addition, most studies focus on extracting sentence-level events and ignore cross-sentence event information. To solve these problems, we simplify the process of event extraction, reduce the processing steps, and combine the two sub-tasks of relation extraction and argument combination as one sub-task. In addition, we consider document-level event extraction, which not only extracts cross-sentence events but also considers broader context information. Experimental results indicate that our novel approach outperforms prior studies. Additionally, the document-level event extraction model attains the top performance on the BioNLP’11 test data and achieves near-leading performance on the BioNLP’13 test data.}, archive = {J_EAAI}, author = {Fangfang Su and Yue Zhang and Pengfei Jiao and Zhidong Zhao and Bobo Li and Fei Li and Donghong Ji}, doi = {10.1016/j.engappai.2025.112001}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {112001}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A two-stage model for unified sentence- and document-level biomedical event extraction}, volume = {161}, year = {2025}, }
• (2025). A position-aware sets based weakly supervised framework for whole-slide subtype classification. EAAI, 161, 111998. (www)

  @article{ , abstract = {Identifying cancer subtypes is essential for personalized treatment and accurate prognosis due to the varying sensitivities of subtypes to therapies. However, in cancer subtype classification tasks, normal slides are usually scarce or absent as negative samples, while subtype whole-slide images (WSIs) often contain extensive unannotated normal tissue regions. These regions introduce significant noise during feature fusion and subtype classification, leading to degraded performance of existing weakly supervised methods In this paper, we propose the Position-aware Sets based Weakly Supervised learning framework (PSWS), designed for cancer subtype classification using WSIs, with a two-stage structure to enhance model efficiency. Specifically, it first presents a novel patch organization approach, distinct from the bag concept of traditional Multiple Instance Learning (MIL), called position-aware sets, as basic units for learning. Then, PSWS automatically selects subtype-specific features based on enhanced histological features and mutual-patch relations, mitigating the negative impact of unannotated negative regions. In the experiments, the superior performance of PSWS over representative MILs is validated through subtype classification tasks on both public datasets and our internally constructed dataset. Furthermore, class probabilities of position-aware sets and attention region visualizations demonstrate its post-hoc interpretability, assisting pathologists in locating suspicious areas.}, archive = {J_EAAI}, author = {Jiuman Song and Bo Yu and Xiaomin Liu and Lele Cong and Zilong Zhou and Xianling Cong and Hongyan Sun and Shuchao Pang and Hechang Chen}, doi = {10.1016/j.engappai.2025.111998}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111998}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A position-aware sets based weakly supervised framework for whole-slide subtype classification}, volume = {161}, year = {2025}, }
• (2025). Action understanding in low-light and pitch-dark conditions: A comprehensive survey. EAAI, 161, 111996. (www)

  @article{ , abstract = {Action understanding in low-light and dark environments (AULLD) is a critical and technically demanding challenge at the intersection of computer vision and engineering applications. This task focuses on interpreting human actions using motion cues under visually constrained conditions, such as poor illumination or complete darkness. Since 2010, machine learning and, more notably, deep learning have significantly advanced AULLD by enabling systems to extract and learn discriminative features from complex, low-visibility data. These techniques have resulted in state-of-the-art solutions, with recent methods achieving accuracies on various benchmark datasets of up to 97.10%, and a 73.20% top-1 accuracy on an infrared dataset. These solutions have had applications across domains such as healthcare monitoring, surveillance and security, human–computer interaction, and social computing. In this survey, we provide the first comprehensive overview of the progress of the research and topical developments in AULLD. We cover a broad range of elements, including the datasets, evaluation protocols, methods, challenges, and emerging research directions. To organize and evaluate the literature in a structured, integrated way, we introduce a novel taxonomic framework, grounded in four core dimensions: body representation, temporal representation, feature representation, and neural architectures. Employing this taxonomy, we perform an analysis of AULLD approaches, exploring their methodological foundations, system design considerations, performance outcomes, and inherent trade-offs. The survey concludes with an in-depth discussion of the existing challenges confronting AULLD, highlighting several promising research directions that hold potential for advancing the field and setting the stage for future innovations and improvements in AULLD.}, archive = {J_EAAI}, author = {Muhammad Munsif and Samee Ullah Khan and Noman Khan and Altaf Hussain and Sung Wook Baik}, doi = {10.1016/j.engappai.2025.111996}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111996}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Action understanding in low-light and pitch-dark conditions: A comprehensive survey}, volume = {161}, year = {2025}, }
• (2025). Transfer learning framework integrating attention mechanism and domain adaptation for low earth orbit satellite network traffic prediction. EAAI, 161, 111992. (www)

  @article{ , abstract = {Traffic prediction is a crucial prerequisite for planning and even network security in Low Earth Orbit (LEO) satellite networks (LSNs). This paper designed a transfer learning framework for LSN traffic prediction that leverages attention mechanism and domain adaptation. Firstly, by integrating five-dimensional data including global population distribution, local time coefficient, the Internet penetration rate of the country, daily data volume of a single Internet user, and global aeronautical traffic demand, a traffic model that could characterize the traffic situation of the area covered by LEO satellite within a specific time range was constructed. Considering the problem of insufficient online traffic data, knowledge was transferred from terrestrial network traffic (source domain) to satellite network traffic (target domain) by incorporating the Domain-Adversarial Neural Network (DANN) method to tackle the data distribution discrepancies between the source and target domains. Finally, by combining DANN with the attention mechanism, the domain-invariant features of the source domain and the target domain were extracted to predict satellite network traffic accurately. Experimental results show that compared to baseline models, the error of the proposed framework in terms of root mean square error measurement is reduced by 9.57% to 33.47% and 18.85% to 38.99% in the two simulated LSN traffic scenarios. Moreover, this framework has low computational complexity among other transfer learning models, which can lay a foundation for subsequent satellite traffic planning and network security.}, archive = {J_EAAI}, author = {Yan Zhang and Yong Wang and Qingsong Zhao and Yadi Zhai and Zhi Lin and Luda Zhao and Yihua Hu}, doi = {10.1016/j.engappai.2025.111992}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111992}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Transfer learning framework integrating attention mechanism and domain adaptation for low earth orbit satellite network traffic prediction}, volume = {161}, year = {2025}, }
• (2025). Skeleton-based multi-person action recognition towards real-world violence detection. EAAI, 161, 111987. (www)

  @article{ , abstract = {Human action recognition is crucial in intelligent systems like robotics, healthcare, and surveillance, gaining significant attention with the rise of AI in computer vision. This work presents the Multi-Skeleton Action Recognizer (MSAR), a comprehensive multi-stage process specifically designed to enhance the recognition and classification of violent activities involving multiple individuals. The approach aims to improve accuracy and robustness in real violence detection and classification leveraging skeletal data. The proposed MSAR pipeline consists of five main stages: Human detection, Human tracking, Skeleton extraction, Data pre-processing and Padding, and Action classification. Experimental results demonstrate the effectiveness of our multi-stage process, showing significant improvements in multi-person action recognition in violence detection. In the human detection stage, our integration of multiple detection models provides a comprehensive visual comparison of model performance, offering empirical insights into accuracy, computational efficiency, and processing speed. The fusion Bidirectional Long Short-Term Memory - Gated Recurrent Unit (BiLSTM-GRU) architecture designed for action classification integrating Deep Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) components, demonstrated maximum accuracy of 96.46% and an average accuracy of 91.18%. The proposed architecture is explored to potentially enhance the model’s ability to capture temporal dependencies in action sequences. Additionally, our proposed multi-stage pipeline for action recognition exemplifies a modular, systematic approach, where each stage functions as an independent module is expanded, optimized, replaced without affecting others. This modularity enhances the flexibility, adaptability, scalability of the pipeline, providing a solid foundation for future advancements across various tasks, application domains.}, archive = {J_EAAI}, author = {Minh-Trieu Truong and Van-Dung Hoang}, doi = {10.1016/j.engappai.2025.111987}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111987}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Skeleton-based multi-person action recognition towards real-world violence detection}, volume = {161}, year = {2025}, }
• (2025). Malware attack and defense game in fractional-order internet of underwater things: Model-based and model-free approaches. EAAI, 161, 111970. (www)

  @article{ , abstract = {With the rapid development of the Internet of Underwater Things (IoUT), its cybersecurity faces increasingly severe challenges, as malware attacks have become a common and highly destructive threat. However, most existing studies rely on integer-order models or static defense mechanisms, which fail to capture the memory and hereditary properties of underwater malware propagation and lack adaptability in dynamic adversarial environments. To address this issue, this paper constructs a fractional-order propagation model and introduces differential game theory to investigate the dynamic interactions between malware and system defense, aiming to derive the optimal dynamic strategies and the corresponding Nash equilibrium for both parties. Furthermore, two innovative artificial intelligence strategy learning approaches are proposed: a model-based method and a model-free method, designed to approximate the optimal control strategies in the game. Extensive comparative experiments and simulation results demonstrate the effectiveness of the proposed methods in strategy learning. We further discuss the practical applications of these methods in IoUT as well as their scalability to other domains.}, archive = {J_EAAI}, author = {Guiyun Liu and Zulong Peng and Tingting Tan and Xiaojing Zhong and Zhongwei Liang}, doi = {10.1016/j.engappai.2025.111970}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111970}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Malware attack and defense game in fractional-order internet of underwater things: Model-based and model-free approaches}, volume = {161}, year = {2025}, }
• (2025). Semi-supervised graph contrastive learning for emotion recognition based on electroencephalogram signals. EAAI, 161, 111969. (www)

  @article{ , abstract = {Emotion recognition plays a crucial role in affective computing, aiming to identify and interpret human emotions. Since electroencephalography (EEG) measures brain electrical activity associated with emotional processing, its use in emotion recognition has garnered significant attention. However, due to the complexity of EEG, capturing the low-dimensional manifold structure from a high-dimensional EEG signal remains a challenge. Moreover, even though EEG labeling is expensive and requires specialization, the development of emotion recognition models has predominantly relied on labeled EEG datasets. To address these issues, we propose a semi-supervised graph contrastive learning (SGCL) model for EEG-based emotion recognition, leveraging the abundance of unlabeled EEG data to generate clearer representations. The proposed SGCL extracts two frequency-domain features from the EEG signal, differential entropy (DE) and power spectral density (PSD), which are integrated through a symmetric similarity network fusion (SSNF) with graph contrastive learning (GCL) to improve the generalization and representation capability of the graph convolutional network (GCN) to complex EEG data in transductive learning tasks. Extensive experimental results on three public datasets reveal that the proposed SGCL consistently outperforms state-of-the-art models. Notably, on the Shanghai Jiao Tong University EEG Emotion dataset (SEED), SGCL achieves an average accuracy of 99.99% using only 3.5% of the data as labeled input, setting a new benchmark in both effectiveness and efficiency. Further analyses confirm our model learns highly discriminative representations and offers insightful explanations, demonstrating its potential for robust emotion recognition.}, archive = {J_EAAI}, author = {Dae Hyeon Kim and Young-Seok Choi}, doi = {10.1016/j.engappai.2025.111969}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111969}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Semi-supervised graph contrastive learning for emotion recognition based on electroencephalogram signals}, volume = {161}, year = {2025}, }
• (2025). Explainable automated wild-orchid identification combining deep neural networks and bayesian networks. EAAI, 161, 111961. (www)

  @article{ , abstract = {Deep learning has been shown repeatedly to be a successful method of obtaining accurate classifiers. This also applies to orchid identification from digital photographs. However, deep neural networks possess the major weakness of lack of explainability, missing the ability to explain the reasons behind a decision. Nevertheless, most current research regarding automated orchid identification applies this blackbox approach. By contrast, in this paper we propose a new method for trustworthy automated orchid identification combining two complementary methods: deep neural networks and feature-based Bayesian networks, where the Bayesian network is also utilized for providing an explanation of the generated solutions. We use other deep neural networks to extract flower characteristics, the features, from the images which are subsequently fed into the Bayesian network as uncertain evidence. When combining the deep neural network and the Bayesian network as an ensemble classifier, both reaching the same conclusion, an accuracy of 89.4% is achieved, the most trustworthy outcome. With a human-in-the-loop ensemble classifier, validation results are even better, yielding an accuracy of 98.1%. Our approach also exploits the taxonomic knowledge represented in the Bayesian network to provide an explanation of the solutions for every case, reinforcing further trust in the method. The result is an explainable user-in-the-loop ensemble classifier. Providing explainability can help build user trust in a system and may play a major role when it is used as a learning aid for new orchid enthusiasts. Finally, the proposed method may be also of value in many fields other than plant determination.}, archive = {J_EAAI}, author = {Diah Harnoni Apriyanti and Luuk J. Spreeuwers and Peter J.F. Lucas}, doi = {10.1016/j.engappai.2025.111961}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111961}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Explainable automated wild-orchid identification combining deep neural networks and bayesian networks}, volume = {161}, year = {2025}, }
• (2025). A safe multi-agent reinforcement learning algorithm using constraint update projection approach. EAAI, 161, 111929. (www)

  @article{ , abstract = {Traditional reinforcement learning has a major limitation, which is that they optimize agent’s policy purely for maximizing rewards. It completely ignore safety considerations. However, in certain critical engineering fields, ensuring safety is of utmost importance, otherwise it can cause incalculable losses. Therefore, this paper proposes a safe Multi-Agent Constrained Update Projection(MACUP) algorithm, which can safely control agents to complete tasks. We solve this problem from the perspective of policy constraint optimization. Firstly, we derive the new bounds of multi-agent policy performance difference based on a tighter general policy performance difference. It contains generalized advantage estimates, and we utilize these bounds as surrogate functions concerning the objective and constraints. Secondly, to address the coordination issue among multiple agents, we employ a multi-agent sequential policy update framework. Finally, we use a projection method to optimize policies, which has low computational complexity and does not require convex approximation of the surrogate function for solving. It can help us reduce errors. Finally, we have validated our algorithm in two different multi-agent safety environments, and the results show that it is able to satisfy safety constraints while achieving higher rewards.}, archive = {J_EAAI}, author = {Yang Liu and Xiang Feng and Huiqun Yu}, doi = {10.1016/j.engappai.2025.111929}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111929}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {A safe multi-agent reinforcement learning algorithm using constraint update projection approach}, volume = {161}, year = {2025}, }
• (2025). Multi-criteria group decision-making method using spherical fuzzy Z-numbers for smart technology revolution in municipal waste management. EAAI, 161, 111928. (www)

  @article{ , abstract = {Due to the new phenomenon of globalization and the fast growing cities, there is drastic pressure on the conventional methods of waste management hence the need for innovative management wastage that is proactive to the theme of environmental and economic challenge of the modern world. This paper aims at investigating the following research question: How has the introduction of smart technologies impacted waste management in the context of a mid-sized city that is in the cross-road of generating more wastes while at the same time, concerns over ecological attainability. They present a new type of Spherical Fuzzy Z-Number Sets in organizational environments dealing with multi criteria group decision making where it possess higher order of uncertainty than conventional fuzzy sets. For handling the multi facility nature of multi criteria group decision making in waste management, we propose the so-called Additive Ratio Assessment method when the attribute weights are unknown. To ensure that the criteria weights are determined objectively in this research, the CRITIC (CRiteria Importance Through Intercriteria Correlation) technique is used. The first part of the study provides a theoretical framework of spherical fuzzy Z-numbers concerning accuracy, scoring functions, and operations. Then we introduce this framework to actual multi criteria group decision making cases in municipal waste management to show that how Spherical Fuzzy Z-Number can facilitate decision-making processes by dealing with the vagueness and fuzziness of stakeholder preferences. The TODIM (Tomada de Decisao Iterativa Multicriterio) technique is used in this to validate and compare for efficiency of the proposed additive ratio assessment method. Besides, this research will not only enhance the theoretical aspect of fuzzy decision making but also present a feasible and effective framework for handling unsound and random decision making problems especially within the context of urban waste management.}, archive = {J_EAAI}, author = {Shahzaib Ashraf and Muhammad Naeem and Chiranjibe Jana and Maria Akram and Gerhard-Wilhelm Weber}, doi = {10.1016/j.engappai.2025.111928}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111928}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {Multi-criteria group decision-making method using spherical fuzzy Z-numbers for smart technology revolution in municipal waste management}, volume = {161}, year = {2025}, }
• (2025). From image processing to artificial intelligence-driven tools: A comprehensive survey on the evolution of feature extraction methods in paintings. EAAI, 161, 111794. (www)

  @article{ , abstract = {This study explores the computational process of converting qualitative elements of paintings, such as shape, color, texture, and line, into quantitative numerical values, which aid in identifying and extracting features from painting images. These identified features are then used to classify paintings based on artist, art style, genre, and art movements. This review employs a systematic literature review methodology, examining approximately 80 research papers to track trends in this field from 2015 to 2025. With the increasing presence of paintings in online media, museums, and galleries, artificial intelligence (AI) plays a significant role in interpreting these subjective elements. This review examines various image processing techniques in conjunction with AI implementations to extract the local and global features of a painting. These methods are combined with AI models, such as deep learning and computer vision algorithms, to improve feature extraction and provide a more thorough and accurate paintings analysis.}, archive = {J_EAAI}, author = {Rekha Sharma and Rishi Gupta and Aditya Sinha}, doi = {10.1016/j.engappai.2025.111794}, journal = {Engineering Applications of Artificial Intelligence}, month = {12}, pages = {111794}, shortjournal = {Eng. Appl. Artif. Intell.}, title = {From image processing to artificial intelligence-driven tools: A comprehensive survey on the evolution of feature extraction methods in paintings}, volume = {161}, year = {2025}, }

EJOR - 49

• (2026). Optimistic and pessimistic approaches for cooperative games. EJOR, 328(2), 725-733. (www)

  @article{ , abstract = {Cooperative game theory explores how to fairly allocate the joint value generated by a group of decision-makers, but its application is compromised by the large number of counterfactuals needed to compute the value of all coalitions, a problem made even more complicated when externalities are present. We provide a theoretical foundation for a simplification used in many applications, in which the value of a coalition is computed assuming that they either select before or after the complement set of agents, providing optimistic and pessimistic values on what a coalition should receive. In a vast set of problems exhibiting what we call feasibility externalities, we show that ensuring a coalition does not receive more than its optimistic value is always at least as difficult as ensuring it receives its pessimistic value. Furthermore, under the presence of negative externalities, we establish the existence of stable allocations that respect these bounds. Finally, we examine well-known optimization-based applications and their corresponding cooperative games to show how our results lead to new insights and allow the derivation of further results from the existing literature.}, archive = {J_EJOR}, author = {Ata Atay and Christian Trudeau}, doi = {10.1016/j.ejor.2025.09.002}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {725-733}, shortjournal = {Eur. J. Oper. Res.}, title = {Optimistic and pessimistic approaches for cooperative games}, volume = {328}, year = {2026}, }
• (2026). Sustainable product development under profit-sharing crowdfunding: An analytical approach to market structure and government policy. EJOR, 328(2), 704-724. (www)

  @article{ , abstract = {This study develops an integrated analytical framework to examine how crowdfunding, market structure, and government policy interact to shape environmentally sustainable product development (SPD). Focusing on profit-sharing and securities-based crowdfunding, we model how investor opportunity costs, risk preferences, platform fees, and regulatory schemes (voluntary vs. mandatory under fiscal vs. non-fiscal policy regimes) influence firm strategies and outcomes across economic, social, and environmental (ESE) dimensions. Firm behavior is analyzed under monopoly and duopoly settings to explore variation in market power and competitive intensity. Findings reveal that voluntary greening can achieve strong ESE outcomes in monopolistic markets with low financial frictions and environmentally aware consumers. In contrast, competitive or uncertain environments often require benchmark-based regulation and fiscal instruments to sustain environmental investments. Two dominant firm profiles emerge: the Voluntary sustainability leader, which performs well under favorable market and investor conditions without policy intervention, and the Policy-driven strategist, which depends on regulatory standards and fiscal tools to overcome competitive pressures and risk constraints. By formalizing investor-entrepreneur interactions and embedding environmental quality as a strategic variable, this research advances the literature on crowdfunding and market-driven sustainability. It provides actionable insights for aligning crowdfunding design and policy frameworks with the broader goals of green innovation and public sustainability.}, archive = {J_EJOR}, author = {Raziyeh Reza-Gharehbagh and Madeleine Pullman}, doi = {10.1016/j.ejor.2025.08.020}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {704-724}, shortjournal = {Eur. J. Oper. Res.}, title = {Sustainable product development under profit-sharing crowdfunding: An analytical approach to market structure and government policy}, volume = {328}, year = {2026}, }
• (2026). Optimal dividend and scale of business strategies with reinsurance and premium pricing for insurance company. EJOR, 328(2), 694-703. (www)

  @article{ , abstract = {This paper investigates the optimal dividend and business scale strategies aimed at maximizing the value of an insurance company. While prior studies typically assume that insurers can only adjust their business scale through reinsurance, this study extends the framework by allowing the insurer to control the premium rate. Under more realistic market assumptions, we examine the joint optimization problem for two common types of reinsurance — proportional and excess-of-loss — across both arbitrage and non-arbitrage scenarios. We derive the optimal strategies for dividends and premium pricing, along with their corresponding value functions. The results show that the insurer should decrease the premium rate and reduce reinsurance coverage as the surplus increases. The optimal dividend policy follows a barrier strategy. Economic interpretations and numerical examples are provided to illustrate the findings.}, archive = {J_EJOR}, author = {Dingjun Yao and Bo Yang and Xin Xu and Youwei Li and Yizhi Wang}, doi = {10.1016/j.ejor.2025.07.039}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {694-703}, shortjournal = {Eur. J. Oper. Res.}, title = {Optimal dividend and scale of business strategies with reinsurance and premium pricing for insurance company}, volume = {328}, year = {2026}, }
• (2026). Anticipating delays in recruitment: Explainable machine learning for the prediction of hard-to-fill online job vacancies. EJOR, 328(2), 680-693. (www)

  @article{ , abstract = {Online job vacancy (OJV) platforms have transformed the labor market by enabling employers to advertise jobs to a wide audience. Particularly in tight labor markets, quickly identifying vacancies likely to suffer prolonged durations is crucial. This study utilizes data from the Flemish public employment service's OJV platform to examine the effectiveness of machine learning in predicting hard-to-fill vacancies. We achieve notable predictive performance with XGBoost in forecasting recruitment delays and demonstrate the importance of capturing non-linear patterns in OJV data. SHAP (SHapley Additive exPlanations) values reveal that the textual content of vacancies and latent company characteristics are key predictors of hiring delays. Counterfactual-SHAP insights provide practical guidance for refining recruitment strategies, enhancing labor market forecasts, and informing targeted policies.}, archive = {J_EJOR}, author = {Wouter Dossche and Sarah Vansteenkiste and Bart Baesens and Wilfried Lemahieu}, doi = {10.1016/j.ejor.2025.06.027}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {680-693}, shortjournal = {Eur. J. Oper. Res.}, title = {Anticipating delays in recruitment: Explainable machine learning for the prediction of hard-to-fill online job vacancies}, volume = {328}, year = {2026}, }
• (2026). Using diet optimization and machine learning for the design of healthy and acceptable menu plans. EJOR, 328(2), 668-679. (www)

  @article{ , abstract = {The success of dietary plans relies on understanding and modelling consumer acceptance, yet quantifying this poses a challenge due to the complexity of individual preferences. Recent research is focused on deriving acceptability constraints directly from data, as demonstrated by its application in designing food baskets with a limited number of commodities. In this study, we applied diet optimization with machine learning to the more complex task of menu planning. This involved considering hundreds of potential food alternatives and assessing their compatibility within a meal using a recipe completion algorithm. Compared to the traditional diet modelling approach of food group filtering, the recipe completion model delivered diets with either higher nutritional adequacy or greater substitute acceptability, depending on the number of food groups used in the traditional method. While more research is needed to further improve the acceptability of substitutions, combining diet optimization with recipe completion presents a promising approach to enhance the nutritional adequacy of individual diets while maintaining the acceptability of food combinations within meals.}, archive = {J_EJOR}, author = {Dominique van Wonderen and Johanna C. Gerdessen and Alida Melse-Boonstra and Marleen C. Onwezen}, doi = {10.1016/j.ejor.2025.06.015}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {668-679}, shortjournal = {Eur. J. Oper. Res.}, title = {Using diet optimization and machine learning for the design of healthy and acceptable menu plans}, volume = {328}, year = {2026}, }
• (2026). Project monitoring and control with an empirically grounded budget-release model. EJOR, 328(2), 646-667. (www)

  @article{ , abstract = {Project monitoring and control (PMC) is a process of measuring a project’s progress and taking corrective action when necessary to ensure successful project completion. However, most existing models lack empirical validation of their assumptions and effectiveness, which limits their practical use. We fill in this gap by using empirical data from 97 real projects to calibrate activity-duration distributions and assess activity-duration dependencies, integrating these empirical foundations into an enhanced PMC model. We further improve the model by incorporating budget-release timing constraints and introducing two new policies for crashing and fast-tracking based on a project’s specific time and cost characteristics. Extensive computational experiments using empirical and artificial data evaluate the effectiveness of these policies. Because the budget-release policies and corrective action types depend on project characteristics such as activity-duration dependencies and topological network structure, key insights from this study can be usefully applied by project managers as heuristics even without a detailed model of their project.}, archive = {J_EJOR}, author = {Jie Song and Jinbo Song and Tyson Browning and Mario Vanhoucke}, doi = {10.1016/j.ejor.2025.06.007}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {646-667}, shortjournal = {Eur. J. Oper. Res.}, title = {Project monitoring and control with an empirically grounded budget-release model}, volume = {328}, year = {2026}, }
• (2026). The continuous roll-on roll-off dual cycling problem with tugs and driver-handled cargo units. EJOR, 328(2), 633-645. (www)

  @article{ , abstract = {Roll-on roll-off vessels are a popular mode of transport in short-sea shipping. In this domain, appropriate stevedoring procedures are crucial to enhance efficiency. This includes dual cycling, where tugs simultaneously load and unload the vessel. Dual cycling reduces turnaround time, thereby giving the vessel more time to travel and allowing for slow steaming and reduced emissions. We extend the roll-on roll-off dual cycling problem by incorporating a continuous time horizon and differentiating cargo units that are handled by their own drivers and units that must be handled by a tug. We propose a mixed integer linear programming model for generating an efficient schedule that minimizes overall makespan by optimizing the sequence of cargo units and the assignment of cargo units to tugs. To solve instances of real-world size with acceptable computational effort, we provide a range of heuristics, including a biased random-key genetic algorithm. Compared to the linear programming model and on instances of real-world size, the genetic algorithm finds good solutions quickly. We derive managerial insights from a sensitivity analysis and show that dual cycling and strategic positioning of driver-handled units can reduce turnaround time by 14.5%, reducing emissions of the considered vessel by more than 8%. We demonstrate the robustness of these insights in uncertain environments through a simulation study.}, archive = {J_EJOR}, author = {Teresa Marquardt and Arne Heinold and Catherine Cleophas and Frank Meisel}, doi = {10.1016/j.ejor.2025.05.050}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {633-645}, shortjournal = {Eur. J. Oper. Res.}, title = {The continuous roll-on roll-off dual cycling problem with tugs and driver-handled cargo units}, volume = {328}, year = {2026}, }
• (2026). Optimizing treatment allocation in the presence of interference. EJOR, 328(2), 620-632. (www)

  @article{ , abstract = {In Influence Maximization (IM), the objective is to — given a budget — select the optimal set of entities in a network to target with a treatment so as to maximize the total effect. For instance, in marketing, the objective is to target the set of customers that maximizes the total response rate, resulting from both direct treatment effects on targeted customers and indirect, spillover, effects that follow from targeting these customers. Recently, new methods to estimate treatment effects in the presence of network interference have been proposed. However, the issue of how to leverage these models to make better treatment allocation decisions has been largely overlooked. Traditionally, in Uplift Modeling (UM), entities are ranked according to estimated treatment effect, and the top entities are allocated treatment. Since, in a network context, entities influence each other, the UM ranking approach will be suboptimal. The problem of finding the optimal treatment allocation in a network setting is NP-hard, and generally has to be solved heuristically. To fill the gap between IM and UM, we propose OTAPI: Optimizing Treatment Allocation in the Presence of Interference to find solutions to the IM problem using treatment effect estimates. OTAPI consists of two steps. First, a causal estimator is trained to predict treatment effects in a network setting. Second, this estimator is leveraged to identify an optimal treatment allocation by integrating it into classic IM algorithms. We demonstrate that this novel method outperforms classic IM and UM approaches on both synthetic and semi-synthetic datasets.}, archive = {J_EJOR}, author = {Daan Caljon and Jente Van Belle and Jeroen Berrevoets and Wouter Verbeke}, doi = {10.1016/j.ejor.2025.09.015}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {620-632}, shortjournal = {Eur. J. Oper. Res.}, title = {Optimizing treatment allocation in the presence of interference}, volume = {328}, year = {2026}, }
• (2026). Soft regression trees: A model variant and a decomposition training algorithm. EJOR, 328(2), 607-619. (www)

  @article{ , abstract = {Decision trees are widely used for classification and regression tasks in a variety of application fields due to their interpretability and good accuracy. During the past decade, growing attention has been devoted to globally optimized decision trees with deterministic or soft splitting rules at branch nodes, which are trained by optimizing the error function over all the tree parameters. In this work, we propose a new variant of soft multivariate regression trees (SRTs) where, for every input vector, the prediction is defined as the linear regression associated to a single leaf node, namely, the leaf node obtained by routing the input vector from the root along the branches with higher probability. SRTs exhibit the conditional computational property, i.e., each prediction depends on a small number of nodes (parameters), and our nonlinear optimization formulation for training them is amenable to decomposition. After showing a universal approximation result for SRTs, we present a decomposition training algorithm including a clustering-based initialization procedure and a heuristic for rerouting the input vectors along the tree. Under mild assumptions, we establish asymptotic convergence guarantees. Experiments on 15 well-known datasets indicate that our SRTs and decomposition algorithm yield higher accuracy and robustness compared with traditional soft regression trees trained using the nonlinear optimization formulation of Blanquero et al. (2021), and a significant reduction in training times as well as a slightly better average accuracy compared with the mixed-integer optimization approach of Bertsimas and Dunn (2019). We also report a comparison with the Random Forest ensemble method.}, archive = {J_EJOR}, author = {Antonio Consolo and Edoardo Amaldi and Andrea Manno}, doi = {10.1016/j.ejor.2025.08.050}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {607-619}, shortjournal = {Eur. J. Oper. Res.}, title = {Soft regression trees: A model variant and a decomposition training algorithm}, volume = {328}, year = {2026}, }
• (2026). An explainable machine learning framework for recurrent event data analysis. EJOR, 328(2), 591-606. (www)

  @article{ , abstract = {This paper introduces a novel explainable temporal point process (TPP) model, Stratified Hawkes Point Process (SHPP), for modelling recurrent event data (RED). Unlike existing approaches that treat temporal influence as a black box or rely on post-hoc explanations, SHPP structurally decomposes event intensities into semantically meaningful components for describing self-, Markovian, and joint influences. This decomposition enables direct quantification of how past events contribute to future event risks, termed as influence values. We further provide a sufficient condition for mean-square stability based on kernel decay, ensuring long-term boundedness of intensities and realistic behavioural predictions. Experiments and an e-commerce case study demonstrate SHPP’s ability to deliver accurate, interpretable, and stable modelling of complex event-driven systems.}, archive = {J_EJOR}, author = {Qi Lyu and Shaomin Wu}, doi = {10.1016/j.ejor.2025.09.005}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {591-606}, shortjournal = {Eur. J. Oper. Res.}, title = {An explainable machine learning framework for recurrent event data analysis}, volume = {328}, year = {2026}, }
• (2026). Decision space dynamic niching-based method for constrained multiobjective evolutionary optimization. EJOR, 328(2), 574-590. (www)

  @article{ , abstract = {Finding a set with a good approximation to the Pareto-optimal solutions in the multiobjective optimization problem (MOP) is a challenging task in terms of convergence toward and diversity across the Pareto optimal front (PoF). In some cases, solving MOPs requires satisfying certain constraints, which significantly increases the complexity of the problem. Such problems are constrained multiobjective optimization problems (CMOPs) and pose considerable computational challenges. Many constrained multiobjective evolutionary algorithms (CMOEAs) face challenges in avoiding becoming trapped in local optima, which impacts convergence, and offer solutions that lack good coverage of the PoF, implying weak diversity. All these nonoptimal or partially optimal solutions in the objective space are essentially clustered in local optimality dilemmas in the decision space. To better eliminate the convergence and diversity challenges caused by clustered solutions, this paper proposes a decision space dynamic niching-based (DSDN) method to better address CMOPs. Specifically, the DSDN method adds a dynamic decision space niche as an additional criterion to the traditional Pareto-constrained dominance principle (Pareto-CDP). The better preserved solutions must satisfy the Pareto-CDP and the condition within the niche radius of other solutions, which strictly meets the original dominance relationship requirement while relaxing the nondominance threshold. As a result, the dynamic adjustment of the niche radius ( N R ) effectively balances the exploitation and exploration of solutions in the decision space while enhancing both convergence and diversity in the objective space. Experiments conducted on four widely recognized test suites and three real-world case studies have demonstrated that the DSDN method yields significantly better results than the original Pareto-CDP algorithms. Furthermore, the proposed approach is competitive with or comparable to seven other state-of-the-art CMOEAs.}, archive = {J_EJOR}, author = {Fan Yu and Qun Chen and Jinlong Zhou}, doi = {10.1016/j.ejor.2025.07.002}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {574-590}, shortjournal = {Eur. J. Oper. Res.}, title = {Decision space dynamic niching-based method for constrained multiobjective evolutionary optimization}, volume = {328}, year = {2026}, }
• (2026). Impact of technology spillovers and subsidies on innovation consortia dynamics. EJOR, 328(2), 560-573. (www)

  @article{ , abstract = {In innovation consortia, governments and enterprises view technology spillovers and subsidies differently, which reduces the effectiveness of policies and undermines cooperation. This work investigates the dynamics of cooperation among leading enterprises (LEs) and small/medium enterprises (SMEs), using an evolutionary game model with intrinsic, extrinsic and mixed strategies. We show that there may exist more than one evolutionary equilibrium of enterprise cooperation under multi-strategy choices. A high subsidy may induce enterprises to speculate and reach a suboptimal evolutionary equilibrium. Furthermore, our findings indicate that SMEs are particularly sensitive to changes in subsidy levels, income distribution ratios, and project income, with their speculative behavior exhibiting a lagging effect. Also, increasing the income distribution ratio of LEs can promote active cooperation between the groups and reduce the speculative tendency of SMEs. This occurs because it magnifies the effect of technology spillover on the income of SMEs and enhances the income of the leading enterprises. Our work improves understanding of cooperative factors and provides new recommendations for government and leading enterprises to improve cooperation and managerial performance.}, archive = {J_EJOR}, author = {Zheng Yang and Lin Li and Nicholas G. Hall and Yongzeng Lai and Yijiang Zhou}, doi = {10.1016/j.ejor.2025.06.031}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {560-573}, shortjournal = {Eur. J. Oper. Res.}, title = {Impact of technology spillovers and subsidies on innovation consortia dynamics}, volume = {328}, year = {2026}, }
• (2026). Online quality endorsement to improve consumer trust: Blockchain or self-hosted livestream?. EJOR, 328(2), 545-559. (www)

  @article{ , abstract = {Online products have long suffered from consumer distrust, putting them at a disadvantage when online brands compete with offline brands. To address this issue, many online brands have adopted blockchain as a means of quality endorsement to improve consumer trust. Alternatively, livestream e-commerce has shown the capability of both quality endorsement and demand creation by real-time interactions with consumers because the application of AR/VR technology and the use of online sales force can effectively induce the herding mentality. For many small and medium-sized online brands, it remains unclear which approach is better, so we formulate the key tradeoffs in the online brand's choice to improve consumer trust in the presence of offline brand's competition. Our research delves into the influence of three key factors on livestream e-commerce: the cost associated with adopting blockchain technology, the strength of positive network externality, and the potential downside of consumer returns. Contrary to conventional wisdom, we find that when the return cost is high and the network externality in livestream is weak, opting for livestream as a quality endorsement can actually benefit the online brand. We also find that the online brand is capable of mitigating the return cost by transferring it to consumers through charging a high retail price, which increases the likelihood of favoring livestream. Our findings shed light on the building and improvement of online consumer trust, contributing to the high-quality development of online-offline business in the new era of consumption.}, archive = {J_EJOR}, author = {Baozhuang Niu and Jian Dong and Xinhu Yu and Yulan Wang}, doi = {10.1016/j.ejor.2025.04.010}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {545-559}, shortjournal = {Eur. J. Oper. Res.}, title = {Online quality endorsement to improve consumer trust: Blockchain or self-hosted livestream?}, volume = {328}, year = {2026}, }
• (2026). Production trade-offs in free disposal hull technologies. EJOR, 328(2), 530-544. (www)

  @article{ , abstract = {In data envelopment analysis, production trade-offs are value judgements that represent simultaneous changes to the inputs and outputs assumed to be technologically possible for any production unit in the technology. The specification of production trade-offs generally leads to an enlargement of the model of technology and increasing its discriminating power on efficiency. In conventional convex variable and constant returns-to-scale models, production trade-offs are the dual forms of weight restrictions. In this paper, we extend the use of production trade-offs to the free disposal hull model of technology and its constant, non-increasing and non-decreasing returns-to-scale variants, in a single unifying development. We provide an axiomatic definition of the new nonconvex technologies, explore the notion of consistent trade-offs in such technologies and develop methods for its testing. We further develop different computational approaches for nonconvex models with production trade-offs. We illustrate the new models by an application in the context of higher education.}, archive = {J_EJOR}, author = {Mahmood Mehdiloo and Grammatoula Papaioannou and Victor V. Podinovski}, doi = {10.1016/j.ejor.2025.06.032}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {530-544}, shortjournal = {Eur. J. Oper. Res.}, title = {Production trade-offs in free disposal hull technologies}, volume = {328}, year = {2026}, }
• (2026). Price optimization for round trip car sharing. EJOR, 328(2), 511-529. (www)

  @article{ , abstract = {Car sharing, car clubs and short-term rentals could support the transition toward net zero but their success depends on them being financially sustainable for service providers and attractive to end users. Dynamic pricing could support this by incentivizing users while balancing supply and demand. We describe the usage of a round trip car sharing fleet by a continuous time Markov chain model, which reduces to a multi-server queuing model where hire duration is assumed independent of the hourly rental price. We present analytical and simulation optimization models that allow the development of dynamic pricing strategies for round trip car sharing systems; in particular identifying the optimal hourly rental price. The analytical tractability of the queuing model enables fast optimization to maximize expected hourly revenue for either a single fare system or a system where the fare depends on the number of cars on hire, while accounting for stochasticity in customer arrival times and durations of hire. Simulation optimization is used to optimize prices where the fare depends on the time of day or hire duration depends on price. We present optimal prices for a given customer population and show how the expected revenue and car availability depend on the customer arrival rate, willingness-to-pay distribution, dependence of the hire duration on price, and size of the customer population. The results provide optimal strategies for pricing of car sharing and inform strategic managerial decisions such as whether to use time- or state-dependent pricing and optimizing the fleet size.}, archive = {J_EJOR}, author = {Christine S.M. Currie and Rym M’Hallah and Beatriz Brito Oliveira}, doi = {10.1016/j.ejor.2025.06.024}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {511-529}, shortjournal = {Eur. J. Oper. Res.}, title = {Price optimization for round trip car sharing}, volume = {328}, year = {2026}, }
• (2026). The development strategy of supply chain intelligent technology considering technology development uncertainty. EJOR, 328(2), 496-510. (www)

  @article{ , abstract = {It’s crucial for both manufacturing and logistics industries to improve logistics efficiency and reduce logistics loss during transportation, storage, and other processes through the development and application of intelligent technology. By focusing on three potential modes of intelligent technology development cooperation between a manufacturer and its logistics provider, we examine the impact of such collaboration on reducing Logistics loss, as well as explore the optimal mode of cooperation for both firms. Our analytical results indicate that compared to independent technology development, collaborative development of intelligent technology can mitigate the adverse effects of double-marginalization. Comparing the three modes of cooperation, we find that higher development cost can incentivize the collaboration between two firms, while higher integration cost and price elasticity may make the cost sharing mode preferable. It is noteworthy that the uncertainty of intelligent technology development exerts a significant moderating effect on the choice of cooperation mode. Heightened technology development uncertainty tends to incentivize both firms to pursue joint development in order to alleviate the negative impact of the uncertainty.}, archive = {J_EJOR}, author = {Peng Han and Yanfang Huo and Weihua Liu and Ershi Qi and Helen Cai}, doi = {10.1016/j.ejor.2025.07.026}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {496-510}, shortjournal = {Eur. J. Oper. Res.}, title = {The development strategy of supply chain intelligent technology considering technology development uncertainty}, volume = {328}, year = {2026}, }
• (2026). Optimal price, quantity, and return policy decisions of a two-period newsvendor with product reviews. EJOR, 328(2), 477-495. (www)

  @article{ , abstract = {Online retailers (e-tailers) have high product return rates. To decrease product returns and consumers’ purchase risk, e-tailers can increase available product information through product reviews and/or offer money back guarantees (MBGs). We examine the rational expectations equilibrium of an e-tailer selling to myopic and strategic consumers over two periods. By the end of period 1, product reviews become available, providing a signal to strategic consumers who may wait to purchase in period 2. The e-tailer decides order quantity, prices, and return policy. We find that under both no returns and MBG, as strategic consumers’ patience increases, the optimal range of having them purchase in period 2 increases. For low signal accuracy, it is optimal to have strategic consumers purchase in period 1, whereas for high signal accuracy, it is optimal to have them purchase in period 2. However, under no returns and for a low signal accuracy, it may be optimal to have strategic consumers purchase in period 2 if their proportion is medium and they are patient. We also find that as consumers’ patience increases, the range where MBG dominates no returns increases. For heterogeneous signal accuracy among strategic consumers, equilibrium strategies are similar to the homogeneous case, except that when consumers’ patience is low, high heterogeneity allows the e-tailer to price discriminate, making low-signal accuracy consumers purchase in period 1 and high-signal accuracy consumers purchase in period 2. Also, as the proportion of low-signal accuracy consumers increases, price discrimination increases. Therefore, heterogeneity increases profit.}, archive = {J_EJOR}, author = {Huirong Fan and Moutaz Khouja and Jing Zhou}, doi = {10.1016/j.ejor.2025.06.023}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {477-495}, shortjournal = {Eur. J. Oper. Res.}, title = {Optimal price, quantity, and return policy decisions of a two-period newsvendor with product reviews}, volume = {328}, year = {2026}, }
• (2026). Probabilistic forecast aggregation with statistical depth. EJOR, 328(2), 460-476. (www)

  @article{ , abstract = {This paper considers aggregation methods for interval forecasts and forecasts of cumulative distribution functions (CDFs) when there are many forecasters, and past forecast accuracy may not be known. For aggregation, the median and trimmed means have been proposed as simple and robust alternatives to the mean, with some trimmed mean approaches enabling recalibration to widen or narrow the resulting interval or CDF forecast. For interval forecast aggregation, the median and trimming are applied to each bound separately. To try to use the available information better, we treat the bounds as a bivariate point with statistical depth used to order the points in terms of centrality. The deepest point can be viewed as the median interval forecast, and the depth of each point can be used as the basis for trimming. For CDF forecasts, the literature presents aggregation methods for which the median or trimmed mean are obtained for each point on the domain of the distribution. However, if one part of a CDF forecast is outlying, the appeal of using the rest of the CDF forecast is perhaps reduced. We use functional depth to provide a measure of centrality for each CDF forecast, and hence identify the deepest function, which can be viewed as the median forecast. We also use functional depth as the basis for trimming, and consider weighted depth to control the width of the resulting aggregated interval or CDF forecast. We provide empirical illustration using data from surveys of professional macroeconomic forecasters, and an application to growth-at-risk.}, archive = {J_EJOR}, author = {James W. Taylor}, doi = {10.1016/j.ejor.2025.06.028}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {460-476}, shortjournal = {Eur. J. Oper. Res.}, title = {Probabilistic forecast aggregation with statistical depth}, volume = {328}, year = {2026}, }
• (2026). A data-driven approach for strategic inventory placement in multi-echelon supply networks. EJOR, 328(2), 446-459. (www)

  @article{ , abstract = {This paper provides a data-driven solution for optimizing inventory buffers in large-scale supply networks. We study the placement and sizing of strategic inventories in multi-echelon supply chains where the decision maker faces uncertain demand with an unknown distribution influenced by explanatory variables. State-of-the-art multi-echelon inventory optimization models, such as the well-known guaranteed-service model (GSM), are non-linear and typically informed by distributional and parametric assumptions. They often rely on dynamic programming and are difficult to solve for large networks. We adapt the GSM to introduce a nonparametric, feature-driven approach to supply chain safety stock optimization that is based on mixed-integer linear programming (MILP). The MILP formulation sets cost-optimal base stocks, which are learned as linear functions of feature data under consideration of service level requirements. This integrated estimation and optimization approach is solved with commercial mathematical programming solvers and is enhanced by a Benders decomposition method for large networks. We extend the literature on data-driven inventory control by a multi-period and multi-echelon approach for safety stock planning in general, acyclic networks. On the real-world networks from Willems (2008), we find that incorporating feature information when setting safety stocks in large supply chains, on average, reduces operational costs out-of-sample. This value of feature information that the proposed model offers to decision-makers increases in demand volatility and is dependent on certain network characteristics.}, archive = {J_EJOR}, author = {Josef Svoboda and Stefan Minner}, doi = {10.1016/j.ejor.2025.06.022}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {446-459}, shortjournal = {Eur. J. Oper. Res.}, title = {A data-driven approach for strategic inventory placement in multi-echelon supply networks}, volume = {328}, year = {2026}, }
• (2026). Minimizing the maximum flow loss in the network maintenance scheduling problem with flexible arc outages. EJOR, 328(2), 430-445. (www)

  @article{ , abstract = {We investigate a network maintenance scheduling problem where maintenance tasks are carried out on the arcs of a network within flexible time windows. During maintenance, an arc is interrupted and no flow can pass through it. Such arc outages introduce flow loss and thus affect network capacity and service capability. For some public services operated on networks, possible blackouts and serious flow loss introducing extreme risks are generally unacceptable. The problem is to find a feasible schedule of maintenance tasks so that the maximum flow loss during the planning horizon is minimized. We introduce a mixed integer programming formulation and a Benders reformulation for the problem. A Benders decomposition algorithm based on a branch-and-cut framework is designed. Strengthened initial cuts and effective cuts are introduced to reduce feasible region so that the exact algorithm is accelerated. An efficient separation procedure is proposed to generate Benders optimality cuts. Computational experiments were conducted on a set of benchmark instances and a set of simulated instances based on telecommunication networks. Computational results show that our algorithm performs much better than applying a solver to the formulation and an existing Benders decomposition algorithm for a related problem. Optimal schedules can reduce extreme risks caused by large flow loss on the network. Since multiple optimal solutions may exist, hierarchical optimization is used to further select a desirable schedule, by either minimizing total flow loss or minimizing the duration of maximum flow loss. With adaptations, our algorithm also performs well for two extensions.}, archive = {J_EJOR}, author = {Shuang Jin and Ying Liu and Jing Zhou and Qian Hu}, doi = {10.1016/j.ejor.2025.07.056}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {430-445}, shortjournal = {Eur. J. Oper. Res.}, title = {Minimizing the maximum flow loss in the network maintenance scheduling problem with flexible arc outages}, volume = {328}, year = {2026}, }
• (2026). Stochastic quay partitioning problem. EJOR, 328(2), 415-429. (www)

  @article{ , abstract = {In this paper we consider the problem of dividing a quay of a container terminal into berth segments so that the quality of service for future ship arrivals is as good as possible. Since future arrivals are unknown, the alternative solutions are evaluated on various arrival scenarios generated for certain arrival intensity from a stochastic model referred to as a ship traffic model (STM). This problem will be referred to as a stochastic quay partitioning problem (SQPP). SQPP is defined by an STM, arrival intensity, quay length and a set of admissible berth lengths. Evaluation of an SQPP solution on one scenario is a problem of scheduling the arriving vessels on the berths, which is a classic berth allocation problem (BAP). In SQPP the sizes of BAP instances that must be solved by far exceed capabilities of the methods presented in the existing literature. Therefore, a novel approach to solving BAP is applied. Tailored portfolios of algorithms capable of solving very large BAP instances under limited runtime are used. Features of SQPP solutions are studied experimentally: patterns in selected berth lengths, dispersion of solutions quality and solutions similarity. We demonstrate, that partitioning a quay into equal-length berths is not the best approach. The largest vessel traffic is dominating in defining best quay partitions, but dedicating quays for shorter vessels give lower dispersion of solution quality. A set of algorithms to partition a quay is proposed and evaluated: methods based on integer linear programming (ILP) to match vessel classes arrival intensities with berth availability, hill climber, tabu search and a greedy approach. Only under high arrival intensity can these methods show their prowess. ILP methods have an advantage of low solution evaluation cost. Tabu is most flexible, but at high evaluation costs. To the best of our knowledge, SQPP is posed and solved for the first time in the operations research.}, archive = {J_EJOR}, author = {Jakub Wawrzyniak and Maciej Drozdowski and Éric Sanlaville}, doi = {10.1016/j.ejor.2025.07.043}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {415-429}, shortjournal = {Eur. J. Oper. Res.}, title = {Stochastic quay partitioning problem}, volume = {328}, year = {2026}, }
• (2026). Scheduling mixed batch machines with inclusive processing set restrictions and non-identical capacities. EJOR, 328(2), 407-414. (www)

  @article{ , abstract = {A new batch scheduling problem, name mixed batch scheduling problem, is received attentions recently. In a mixed batch scheduling model, the processing time of a job batch H is defined as α max j ∈ H { p j } + ( 1 − α ) ∑ j ∈ H p j , where α ∈ [ 0 , 1 ] is a constant. In other words, the processing time of a job batch is the weighted sum of the maximum processing time and the total processing time of jobs in the batch. In this paper, we study the problem of scheduling mixed batch machines with non-identical capacities under inclusive processing set restrictions, where the objective is to minimize the makespan of finishing all the jobs. We present a fast approximation algorithm with a performance ratio of 4 / 3 + α for the problem, which improves up the existing performance bounds in the literature. By providing a technical lemma, we are able to develop the first polynomial time approximation scheme (PTAS) for the problem. We also design linear-time approximation schemes for two important special cases of the problem.}, archive = {J_EJOR}, author = {Jinwen Ou and Weidong Li}, doi = {10.1016/j.ejor.2025.07.012}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {407-414}, shortjournal = {Eur. J. Oper. Res.}, title = {Scheduling mixed batch machines with inclusive processing set restrictions and non-identical capacities}, volume = {328}, year = {2026}, }
• (2026). A matheuristic approach for the robust coloured travelling salesman problem with multiple depots. EJOR, 328(2), 390-406. (www)

  @article{ , abstract = {In this work, a special type of the travelling salesman problem (TSP) called the coloured TSP (CTSP) is considered. The CTSP, which has many real-world applications, involves a set of salesmen, each assigned a specific colour, and cities that may have one or multiple colours. Salesmen are restricted to visiting only cities that share their colour. We consider a specific depot for each salesman, and the edge weights are uncertain, meaning that there is a set of possible scenarios for their values. A robust objective is considered and minimised using an artificial intelligence (AI)-driven matheuristic approach due to the high computational complexity of the problem. This approach integrates a variable neighbourhood search (VNS) framework with genetic algorithm (GA) and simulated annealing (SA) operators. More importantly, local improvements based on mathematical programming are applied to different parts of a proportion of the solutions using the concept of partial optimisation metaheuristic under special intensification conditions (POPMUSIC). A key innovation of our method is the use of an artificial neural network to guide the POPMUSIC procedure by selecting only solution segments with high improvement potential, thereby reducing computation time. Extensive computational experiments demonstrate the effectiveness of the proposed algorithm, which outperforms four state-of-the-art methods in solution quality and runs faster than three of them. We also investigate the contribution of individual algorithmic components and the cost of robustness. Furthermore, our method improves upon the best-known results for the single-depot deterministic version of the CTSP from the literature.}, archive = {J_EJOR}, author = {Abtin Nourmohammadzadeh and Stefan Voß}, doi = {10.1016/j.ejor.2025.06.018}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {390-406}, shortjournal = {Eur. J. Oper. Res.}, title = {A matheuristic approach for the robust coloured travelling salesman problem with multiple depots}, volume = {328}, year = {2026}, }
• (2026). Fifty years of research on resource-constrained project scheduling explored from different perspectives. EJOR, 328(2), 367-389. (www)

  @article{ , abstract = {The resource-constrained project scheduling problem is one of the most investigated problems in the project scheduling literature, and has a rich history. This article provides a perspective on this challenging scheduling problem, without having the ambition to provide a complete overview. Instead, the article does aim to summarize a number of reasons why this problem has been so intensely investigated from different perspectives. It will be shown that this scheduling problem has many faces, and therefore deserves a lot of research time from a computational and theoretical point of view as well as from a practical point of view. An overview of possible extensions to other problems and a detailed overview of the used (both heuristic and exact) solution methods will be given. In addition, the data used will be discussed and interesting avenues for further research will be mentioned throughout the different sections.}, archive = {J_EJOR}, author = {Christian Artigues and Sönke Hartmann and Mario Vanhoucke}, doi = {10.1016/j.ejor.2025.03.024}, journal = {European Journal of Operational Research}, month = {1}, number = {2}, pages = {367-389}, shortjournal = {Eur. J. Oper. Res.}, title = {Fifty years of research on resource-constrained project scheduling explored from different perspectives}, volume = {328}, year = {2026}, }
• (2026). Dynamic mode decomposition for online portfolio selection task. EJOR, 328(1), 349-365. (www)

  @article{ , abstract = {Online portfolio selection (OPS) is a complex task aimed at maximizing investment returns through strategic allocation of capital among risky assets. Traditional Follow the Winner (FTW) strategies, grounded in the Best Constant Rebalanced Portfolios strategy, assume market is independent and identically distributed (i.i.d.), which often fails to capture real-world financial market dynamics, leading to sub-optimal performance in practical applications. To address this limitation, we propose integrating Dynamic Mode Decomposition (DMD) into FTW strategies. DMD is a powerful data-driven technique that originated in the field of fluid dynamics. It is designed to extract coherent structures and identify temporal patterns within complex data. By applying DMD to financial market data, we can uncover underlying patterns and trends that are not apparent under the i.i.d. assumption. Significantly, the integrated DMD in this paper allows for efficient recursion, which is particularly crucial for OPS task. To illustrate the effectiveness of the proposed idea, we consider the Exponential Gradient (EG) strategy as an example and proposed Exponential Gradient with Dynamic Mode Decomposition (EGDMD). The results demonstrate that the proposed EGDMD outperforms traditional EG-type strategies, significantly improves risk-adjusted returns, and maintains computational efficiency. The integration of DMD allows for more accurate identification of market patterns, leading to more effective investment decisions and enhanced portfolio performance.}, archive = {J_EJOR}, author = {Jiahao Li and Yong Zhang and Xiaoteng Zheng}, doi = {10.1016/j.ejor.2025.04.049}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {349-365}, shortjournal = {Eur. J. Oper. Res.}, title = {Dynamic mode decomposition for online portfolio selection task}, volume = {328}, year = {2026}, }
• (2026). Choice-based crowdshipping for next-day delivery services: A dynamic task display problem. EJOR, 328(1), 336-348. (www)

  @article{ , abstract = {This paper studies integrating the crowd workforce into next-day home delivery services. In this setting, both crowd drivers and contract drivers collaborate in making deliveries. Crowd drivers have limited capacity and can choose not to deliver if the presented tasks do not align with their preferences. The central question addressed is: How can the platform minimize the total task fulfilment cost, which includes payouts to crowd drivers and additional payouts to contract drivers for delivering the unselected tasks by customizing task displays to crowd drivers? To tackle this problem, we formulate it as a finite-horizon Stochastic Decision Problem, capturing crowd drivers’ utility-driven task preferences, with the option of not choosing a task based on the displayed options. An inherent challenge is approximating the non-constant marginal cost of serving orders not chosen by crowd drivers, which are then assigned to contract drivers. We address this by leveraging a common approximation technique, dividing the service region into zones. Furthermore, we devise a stochastic look-ahead strategy that tackles the curse of dimensionality issues arising in dynamic task display execution and a non-linear (problem specifically concave) boundary condition associated with the cost of hiring contract drivers. In experiments inspired by Singapore’s geography, we demonstrate that choice-based crowd shipping can reduce next-day delivery fulfilment costs by up to 16.9%. The observed cost savings are closely tied to the task display policies and the task choice behaviours of drivers.}, archive = {J_EJOR}, author = {Alp Arslan and Fırat Kılcı and Shih-Fen Cheng and Archan Misra}, doi = {10.1016/j.ejor.2025.05.046}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {336-348}, shortjournal = {Eur. J. Oper. Res.}, title = {Choice-based crowdshipping for next-day delivery services: A dynamic task display problem}, volume = {328}, year = {2026}, }
• (2026). When cost metafrontiers are nonconvex in the outputs, then the production metafrontier is nonconvex: The price of a convexification strategy. EJOR, 328(1), 324-335. (www)

  @article{ , abstract = {Metafrontier analysis is widely used to account for technological heterogeneity among producers. The approach involves combining a number of group-specific production possibilities sets to form a production possibilities metaset. Even though the union of the group sets normally results in a nonconvex metaset, most authors proceed as if this metaset is convex. Kerstens, O’Donnell and Van de Woestyne (2019) obtain new results on the union operator on sets under various assumptions and empirically illustrate that the popular convexification strategy is highly questionable. In this paper we transpose their results on the union operator from a production to a cost context: this is new. We then explore the extent to which convexity of the cost function is corroborated using a newly developed test. Furthermore, we check to which extent a convexification strategy is tenable when estimating a cost metafrontier. We use an original banking data set from China and the USA to illustrate the main issues. We establish that the cost function is not convex in the outputs for China and that the convexification strategy leads to potentially-biased estimates of the cost metafrontier and associated measures of efficiency.}, archive = {J_EJOR}, author = {Kristiaan Kerstens and Christopher O’Donnell and Ignace Van de Woestyne and Shirong Zhao}, doi = {10.1016/j.ejor.2025.05.048}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {324-335}, shortjournal = {Eur. J. Oper. Res.}, title = {When cost metafrontiers are nonconvex in the outputs, then the production metafrontier is nonconvex: The price of a convexification strategy}, volume = {328}, year = {2026}, }
• (2026). Mitigating adversarial attacks on transformer models in credit scoring. EJOR, 328(1), 309-323. (www)

  @article{ , abstract = {The integration of unstructured data, such as text created by borrowers, offers new opportunities for improving credit default prediction but also introduces new risks. This study examines the robustness of transformer-based credit scoring models that utilize textual data and assesses their vulnerability to adversarial attacks. Using peer-to-peer lending data, we show that small, semantically neutral changes in loan descriptions can substantially alter model outputs. These vulnerabilities expose lenders and borrowers to economic risks through distorted risk assessments and mispriced loans. We evaluate two mitigation strategies: adversarial training and topic modeling. Adversarial training improves robustness without compromising predictive performance. Topic modeling provides a more interpretable and stable representation of borrower narratives. An economic analysis confirms that robust models reduce mispricing and improve outcomes for all parties. The findings underscore the importance of robustness as the use of unstructured data in credit scoring becomes more accessible.}, archive = {J_EJOR}, author = {Brandon Schwab and Johannes Kriebel}, doi = {10.1016/j.ejor.2025.05.029}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {309-323}, shortjournal = {Eur. J. Oper. Res.}, title = {Mitigating adversarial attacks on transformer models in credit scoring}, volume = {328}, year = {2026}, }
• (2026). Cost-sensitive single-index classification model. EJOR, 328(1), 295-308. (www)

  @article{ , abstract = {Single-index models (SIMs) are a type of semiparametric model in which a response variable is assumed to be related to a linear combination of explanatory variables by an unknown function, on which any restriction is imposed. Thus, they provide both interpretability and flexibility to capture complex data relationships. In this paper, SIMs are extended to the cost-sensitive classification problem by minimizing the different misclassification costs. The flexibility of SIMs combined with a cost-sensitive approach results in a powerful model to minimize losses and optimize decision making. This is demonstrated through an extensive simulation study and the analysis of five real data sets, where the proposed approach outperforms both parametric and semi-parametric previous approaches.}, archive = {J_EJOR}, author = {Jorge C-Rella and Ricardo Cao and Juan M. Vilar}, doi = {10.1016/j.ejor.2025.08.058}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {295-308}, shortjournal = {Eur. J. Oper. Res.}, title = {Cost-sensitive single-index classification model}, volume = {328}, year = {2026}, }
• (2026). Diverse ensemble cost-sensitive logistic regression. EJOR, 328(1), 282-294. (www)

  @article{ , abstract = {In recent years, cost-sensitive methods have become increasingly crucial for decision-making in various real-world applications. These methods have been developed for the purpose of minimizing costs or risks for stakeholders. Moreover, the interpretability of cost-sensitive methods has gained considerable attention in critical domains such as finance and medical care. In this article, we propose a diverse ensemble of cost-sensitive logistic regression models to reduce costs for binary classification tasks, as well as a novel algorithm based on the partial conservative convex separable quadratic approximation to solve this non-convex optimization problem. The proposed method demonstrates substantial cost savings through extensive simulations and real-world applications, including fraud detection and gene expression analysis. Additionally, unlike other ensembling techniques, the resulting model of the proposed method is fully interpretable as a logistic regression model and achieves a high level of sparsity induced by the proposed algorithm. We believe this approach offers deeper insights into the relationship between predictors and response, enabling more informed decision-making in practical scenarios.}, archive = {J_EJOR}, author = {Bing Yang and Stefan Van Aelst and Tim Verdonck}, doi = {10.1016/j.ejor.2025.07.028}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {282-294}, shortjournal = {Eur. J. Oper. Res.}, title = {Diverse ensemble cost-sensitive logistic regression}, volume = {328}, year = {2026}, }
• (2026). Multi-view ensemble feature selection via SemiDefinite programming. EJOR, 328(1), 269-281. (www)

  @article{ , abstract = {Multi-view learning faces significant challenges in selecting discriminative features while managing redundancy and noise across heterogeneous data sources. To address these issues, this paper introduces Multi-view Ensemble Feature Selection (MEFS), a novel framework that systematically integrates view generation (VG) and view selection (VS) through a unified optimization paradigm. By reformulating feature selection as a MaxCut problem and leveraging SemiDefinite Programming (SDP) relaxation, MEFS dynamically balances the generalization capability of individual views with their pairwise diversity, eliminating the need for manual parameter tuning. A key innovation is the proposed pairwise diversity metric, which quantifies inter-view dissimilarity using between-class scatter matrices to ensure complementary feature subsets. Extensive experiments on ten benchmark datasets demonstrate that MEFS consistently outperforms state-of-the-art methods in accuracy, robustness, and computational efficiency. Ablation studies validate the synergistic effect of combining VG and VS modules.}, archive = {J_EJOR}, author = {Xiaojian Ding and Xin Wang and Pengcheng Shi}, doi = {10.1016/j.ejor.2025.07.014}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {269-281}, shortjournal = {Eur. J. Oper. Res.}, title = {Multi-view ensemble feature selection via SemiDefinite programming}, volume = {328}, year = {2026}, }
• (2026). On the price of diversity for multiwinner elections under (weakly) separable scoring rules. EJOR, 328(1), 258-268. (www)

  @article{ , abstract = {We consider a model of multi-winner elections, where each voter expresses a linear preference over a finite set of alternatives. Based on voters’ preferences, the primary goal is to select a subset of admissible alternatives, forming what is referred to as a committee. We explore (weakly) separable committee scoring rules, the voting mechanisms that assess each alternative individually using a scoring vector and select the top k alternatives, where k represents the committee’s size. Furthermore, we operate under the assumption that alternatives are categorized based on specific attributes. Within each attribute category, there exists a targeted minimum number of alternatives that the selected committee should encompass, emphasizing the necessity for diversity. In this context, we assess the cost associated with imposing such a diversity constraint on the voting process. This assessment is conducted through two methodologies, referred to as the “price of diversity” and the “individual price of diversity”. We set the upper bounds for both prices across all (weakly) separable committee scoring rules. Additionally, we show how the maximum price of diversity can be used to discriminate between different voting rules in this context. Ultimately, we illustrate that concentrating on the candidates’ performance yields a more accurate estimation of the price of diversity compared to a focus on the enforced diversity constraint.}, archive = {J_EJOR}, author = {Mostapha Diss and Clinton Gubong Gassi and Eric Kamwa}, doi = {10.1016/j.ejor.2025.06.013}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {258-268}, shortjournal = {Eur. J. Oper. Res.}, title = {On the price of diversity for multiwinner elections under (weakly) separable scoring rules}, volume = {328}, year = {2026}, }
• (2026). Nonconvex truncated conditional value at risk-based sparse linear regression. EJOR, 328(1), 246-257. (www)

  @article{ , abstract = {Conditional value at risk (CVaR) is a widely recognized risk measure used to manage data uncertainty within risk management. In this paper, we study a class of sparse linear regression models based on truncated CVaR measure and ℓ 0 -norm regularization. Due to the nonconvexity and nonsmoothness of the objective functions, as well as the NP-hardness of the problem with the ℓ 0 -norm regularization, we propose an approximation model that employs a tight relaxation of the ℓ 0 -norm. The solution equivalence between the proposed model and its approximation model is explored. To efficiently solve the approximation model, we develop a semismooth Newton-based proximal majorization-minimization algorithm. Furthermore, the convergence analysis of the proposed algorithm is presented, and the convergence rate for the reduced CVaR-based sparse linear regression model is established. Moreover, extensive numerical experiments conducted on both synthetic and real datasets validate the stability and effectiveness of the proposed algorithm, demonstrating significant improvements in both sparsity and accuracy compared to existing state-of-the-art methods.}, archive = {J_EJOR}, author = {Boyi Xie and Zhongming Wu and Min Li}, doi = {10.1016/j.ejor.2025.06.004}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {246-257}, shortjournal = {Eur. J. Oper. Res.}, title = {Nonconvex truncated conditional value at risk-based sparse linear regression}, volume = {328}, year = {2026}, }
• (2026). Designing and computing explanations for comparisons inferred from an additive value model. EJOR, 328(1), 232-245. (www)

  @article{ , abstract = {Many decision models are based on an additive representation of preferences. Recommendations obtained from such additive decision models are sometimes considered as self-evident. On the contrary, we claim that these recommendations deserve an explanation so as to be fully understood by the user/decision-maker and to foster her trust. We propose to explain a preference statement x preferred to y by decomposing this statement into simpler ones. Arguments in favor of x (Pros), and arguments in favor of y (Cons) are decomposed using a covering scheme in which each Con is covered by a Pro. We use a decomposition language in which elementary self-evident statements involve ( i ) one Pro against one Con, ( ii ) one pro against several Cons, or ( iii ) several Pros against one Con. We prove that computing such explanations is computationally difficult in case ( ii ) and ( iii ), and propose a mathematical programming formulation to solve it. Numerical experiments provide insights on the actual behavior of our algorithm. We also illustrate the usefulness of our approach in the context of multicriteria decision aid but also for machine learning approaches.}, archive = {J_EJOR}, author = {Manuel Amoussou and Khaled Belahcene and Nicolas Maudet and Vincent Mousseau and Wassila Ouerdane}, doi = {10.1016/j.ejor.2025.05.058}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {232-245}, shortjournal = {Eur. J. Oper. Res.}, title = {Designing and computing explanations for comparisons inferred from an additive value model}, volume = {328}, year = {2026}, }
• (2026). Bayesian variable selection in kriging metamodeling for quality design. EJOR, 328(1), 216-231. (www)

  @article{ , abstract = {The increasing complexity of production processes and rapid developments in digital technology have fueled the adoption of metamodels in quality design. Kriging has emerged as one of the most popular emulation methods for both deterministic and stochastic simulations. Conventional Kriging models with predetermined mean functions, such as ordinary or universal Kriging, may exhibit subpar predictive performance when strong trends exist. This paper proposes a novel variable selection procedure for the mean function that ensures prediction accuracy while using only a limited number of variables to capture the potential existing trends in deterministic simulations. The proposed method integrates the benefits of Bayesian variable selection and frequentist statistical tests. Initially, a group of potential models is chosen to build the mean function, employing the Bayesian method with priors designed to guarantee sparsity. This results in a significant reduction in the number of models to be considered in the next stage. Subsequently, each candidate model undergoes rigorous frequentist tests to thoroughly assess its reliability and validity. Extensive simulation studies are conducted using the well-known Borehole function and a real-life case. The results demonstrate the superiority of the proposed method over several existing approaches, establishing its effectiveness in achieving robust parameter design.}, archive = {J_EJOR}, author = {Baoping Tao and Zifei Han and Wen Shi and Min Wang and Linhan Ouyang}, doi = {10.1016/j.ejor.2025.06.003}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {216-231}, shortjournal = {Eur. J. Oper. Res.}, title = {Bayesian variable selection in kriging metamodeling for quality design}, volume = {328}, year = {2026}, }
• (2026). Design and pricing of extended warranty menus with reference effects. EJOR, 328(1), 201-215. (www)

  @article{ , abstract = {Consumer durables increasingly come with extended warranty menus—beyond manufacturers’ base warranties—that offer multiple options with differentiated protection lengths and prices. When choosing from an extended warranty menu, consumers might not only evaluate the intrinsic utility of each option but also form a reference point against which the available options are compared. In this paper, we posit that consumers compare the price-length ratios of available options when determining their willingness to pay, and investigate the design and pricing of extended warranty menus under such reference effects. To this end, we adapt the standard multinomial logit choice model to incorporate two types of reference point that are generated endogenously and exogenously, respectively, with respect to the given menu. We show that if the warranty options are ordered in the protection length, then the optimal pricing policies prescribe the same adjusted markup and the same price-length ratio alternately. We further extend our analysis to price competition under reference effects, where multiple firms compete in the aftermarket, each offering a single extended warranty. We prove the existence of a unique Nash equilibrium and develop an efficient method to identify it. Numerical examples are presented to illustrate the analytical findings, and sensitivity analyses are conducted to examine the impact of reference-effect coefficients on the optimal pricing policies. Overall, this work highlights the importance of incorporating reference effects into the design and pricing of extended warranty menus.}, archive = {J_EJOR}, author = {Xiao-Lin Wang and Chenglong Li and Junjie Wang}, doi = {10.1016/j.ejor.2025.05.056}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {201-215}, shortjournal = {Eur. J. Oper. Res.}, title = {Design and pricing of extended warranty menus with reference effects}, volume = {328}, year = {2026}, }
• (2026). Planning bayesian reliability demonstration tests via a generalized test statistic. EJOR, 328(1), 189-200. (www)

  @article{ , abstract = {Reliability demonstration testing (RDT) has been extensively employed to verify whether a product meets specific reliability requirements at a desired confidence level. Driven by intense market competition and constrained test resources, manufacturers are motivated to seek effective strategies to reduce the testing efforts required for RDT. In this paper, we propose a method that utilizes existing knowledge and information obtained from the product design and development phase to construct a Bayesian prior distribution of the product’s reliability. Based on this prior, a preliminary disposition decision on whether to accept or reject the product is made. A subsequent demonstration test is needed only when the prior information is deemed insufficient for an immediate disposition. A RDT planning method is developed based on the posterior distribution of the product’s reliability, which is applicable to general cases involving non-conjugate priors. We study two types of demonstration testing: binomial and exponential. For each, we prove the existence of an optimal test plan and develop an efficient searching algorithm to determine it. Numerical studies are conducted to demonstrate the effectiveness of the proposed method, supplemented by a case study on RDT for systems of different configurations. Overall, this work provides a unified and effective framework for reliability demonstration under the Bayesian paradigm.}, archive = {J_EJOR}, author = {Zan Li and Jianyu Xu and Chengjie Wang and Xiao-Lin Wang}, doi = {10.1016/j.ejor.2025.08.011}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {189-200}, shortjournal = {Eur. J. Oper. Res.}, title = {Planning bayesian reliability demonstration tests via a generalized test statistic}, volume = {328}, year = {2026}, }
• (2026). A primal–dual policy iteration algorithm for constrained markov decision processes. EJOR, 328(1), 174-188. (www)

  @article{ , abstract = {The solution algorithms of Constrained Markov Decision Process (CMDP), a widely adopted model for sequential decision-making, have been intensively studied in the literature. Despite increasing effort, the Linear Programming (LP) formulation of CMDP remains the dominant exact method that leads to the optimal solution without constraint violations. However, the LP formulation is computationally inefficient due to the curse of dimensionality in CMDP state and action spaces. In this study, we introduce a novel policy iteration method for CMDP, based on decomposition and row-generation techniques. We design a Primal–Dual Policy Iteration (PDPI) algorithm that utilizes state values and Lagrangian multipliers to improve randomized stationary policies in an iterative fashion. We analytically show that upon convergence, PDPI produces the optimal solution for CMDP. An upper bound of the convergence iterations is also given. To validate the algorithm performance, we conduct comprehensive computational experiments on six benchmarking problems curated from the literature. Results show that PDPI outperforms conventional methods considerably, improving the total algorithm runtime by up to 89.19%. The improvement becomes more significant as the problem size grows larger. We further provide insights and discuss the impact of the developed method.}, archive = {J_EJOR}, author = {Zeyu Liu and Xueping Li and Anahita Khojandi}, doi = {10.1016/j.ejor.2025.08.038}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {174-188}, shortjournal = {Eur. J. Oper. Res.}, title = {A primal–dual policy iteration algorithm for constrained markov decision processes}, volume = {328}, year = {2026}, }
• (2026). Controlling antithetic variates. EJOR, 328(1), 162-173. (www)

  @article{ , abstract = {We establish and investigate a theoretical framework for controlling covariance matrices in the method of antithetic variates through control variates to further reduce estimator variance. Instead of preemptively and carefully designing an estimator vector with negatively correlated components, the proposed framework starts with a predefined estimator vector that incorporates specified control variates. The weights and control matrix are then analytically determined through matrix algebra. The joint optimality of the resulting estimator variance is ensured with respect to both the weights and the control matrix, with closed-form implementable formulas derived for the optimal parameter pair. Numerical results are provided for various typical examples to illustrate the effectiveness, potential, and challenges of the proposed framework.}, archive = {J_EJOR}, author = {Reiichiro Kawai}, doi = {10.1016/j.ejor.2025.08.027}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {162-173}, shortjournal = {Eur. J. Oper. Res.}, title = {Controlling antithetic variates}, volume = {328}, year = {2026}, }
• (2026). Robust multi-period blood inventory routing under multiple uncertainties. EJOR, 328(1), 137-161. (www)

  @article{ , abstract = {We study a multi-period blood inventory routing problem that integrates production, inventory, and distribution decisions under uncertainties in demand, donation supply, and travel times, all while accounting for the limited shelf life of blood products. Our model captures transportation efficiency through a disutility measure based on vehicles’ arrival times at hospitals, and addresses supply–demand imbalances by allowing selective rejection of service requests at a high penalty cost. We formulate a robust optimization model that simultaneously determines production quantities, inventory levels, hospital service selections, and vehicle routing for each period. The objective is to minimize the total cost over the planning horizon, which includes worst-case inventory holding, wastage, and transportation costs, unserved demand penalties, and overall transportation disutility. To obtain an exact solution, we propose an integrated algorithm within the L -shaped framework that combines Benders decomposition with a branch-and-price-and-cut (BPC) scheme. This approach decomposes the robust model into a master problem and period-specific subproblems. For a given master solution, we first use constraint programming to verify the feasibility of the subproblems, and then, if feasible, solve them with a tailored BPC algorithm to generate Benders cuts that eliminate suboptimal master solutions. Extensive numerical experiments, including a case study at the Blood Center in Chongqing, demonstrate the effectiveness of our approach. Our analysis quantifies the benefits of incorporating uncertainty and robustness while providing managerial insights through a systematic evaluation of various parameters.}, archive = {J_EJOR}, author = {Ling Qing and Yunqiang Yin and Joshua Ignatius and Dujuan Wang}, doi = {10.1016/j.ejor.2025.05.036}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {137-161}, shortjournal = {Eur. J. Oper. Res.}, title = {Robust multi-period blood inventory routing under multiple uncertainties}, volume = {328}, year = {2026}, }
• (2026). Data-driven joint optimization of maintenance and spare parts provisioning: A distributionally robust approach. EJOR, 328(1), 122-136. (www)

  @article{ , abstract = {This paper investigates the joint optimization of condition-based maintenance and spare provisioning, incorporating insights obtained from sensor data. Prognostic models estimate components’ remaining lifetime distributions (RLDs), which are integrated into an optimization model to coordinate maintenance and spare provisioning. The existing literature addressing this problem assumes that prognostic models provide accurate estimates of RLDs, thereby allowing a direct adoption of Stochastic Programming or Markov Decision Process methodologies. Nevertheless, this assumption often does not hold in practice since the estimated distributions can be inaccurate due to noisy sensors or scarcity of training data. To tackle this issue, we develop a Distributionally Robust Chance Constrained (DRCC) formulation considering general discrepancy-based ambiguity sets that capture potential distribution perturbations of the estimated RLDs. The proposed formulation admits a Mixed-Integer Linear Programming (MILP) reformulation, where explicit formulas are provided to simplify the general discrepancy-based ambiguity sets. Finally, for the numerical illustration, we test a type- ∞ Wasserstein ambiguity set and derive closed-form expressions for the parameters of the MILP reformulation. The efficacy of our methodology is showcased in a wind turbine case study, where the proposed DRCC formulation outperforms other benchmarks based on stochastic programming and robust optimization.}, archive = {J_EJOR}, author = {Heraldo Rozas and Weijun Xie and Nagi Gebraeel and Stephen Robinson}, doi = {10.1016/j.ejor.2025.06.025}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {122-136}, shortjournal = {Eur. J. Oper. Res.}, title = {Data-driven joint optimization of maintenance and spare parts provisioning: A distributionally robust approach}, volume = {328}, year = {2026}, }
• (2026). Towards blockchain-enabled circular closed-loop supply chain and impact of consumers’ distrust in price, product greenness sensitivity and carbon tax and subsidy. EJOR, 328(1), 105-121. (www)

  @article{ , abstract = {With the increasing emphasis on environmental sustainability, both governments and consumers are more concerned than ever about the greenness of products. In this complex landscape, Supply Chains (SCs) face challenges in building trust and avoiding greenwashing accusations. Blockchain technology offers a promising solution by ensuring transparency and circularity within SCs, particularly in identifying customers for product recycling. This study pioneers the exploration of consumers' distrust in pricing and product greenness, alongside the impact of carbon policies (taxes and subsidies) within a closed-loop supply chain (CLSC). Using classical Stackelberg game theory, we develop two models that identify equilibrium decisions for SC members, focusing on pricing, green production investment, circularity, and blockchain adoption. Additionally, we propose an evolutionary game theory model to find the optimal government policies and identify the long-term behaviour of the CLSC and government in two heterogeneous populations. Our findings reveal that if the retailer's share of blockchain costs falls below a certain threshold, blockchain adoption becomes less profitable than exclusive investment in green production. A higher (lower) subsidy rate benefits (harms) the retailer but disadvantages (benefits) the collector. Blockchain adoption is generally more profitable for manufacturers and retailers, though less so for collectors, and it also drives greater investment in green production. While subsidies encourage blockchain adoption, they are not a sustainable long-term strategy for governments. Ultimately, the evolutionarily stable strategy for SCs involves a balanced investment in both green production and blockchain or green production alone, depending on market characteristics and cost-sharing structures.}, archive = {J_EJOR}, author = {Mohammad Akbarzadeh Sarabi and Ata Allah Taleizadeh and Arijit Bhattacharya}, doi = {10.1016/j.ejor.2025.06.030}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {105-121}, shortjournal = {Eur. J. Oper. Res.}, title = {Towards blockchain-enabled circular closed-loop supply chain and impact of consumers’ distrust in price, product greenness sensitivity and carbon tax and subsidy}, volume = {328}, year = {2026}, }
• (2026). A split-embedded metaheuristic for the heterogeneous inventory routing problem with batch size. EJOR, 328(1), 91-104. (www)

  @article{ , abstract = {Handling transportation and inventory management simultaneously is a challenging problem. The most famous optimization problem in this domain, known as the Inventory Routing Problem (IRP), aims to determine the routes and quantities to be delivered by a set of vehicles to meet customer demands at a minimum total inventory and transportation cost. The vast majority of works carried out so far on the IRP consider a homogeneous fleet of vehicles. This paper addresses, instead, a new IRP variant that considers intrinsic characteristics of real supply chains, such as a period-dependent heterogeneous fleet of vehicles and batch sizes for the delivered quantities. We model the problem with a Mixed Integer Linear Programming (MILP) formulation and propose a Split-Embedded Metaheuristic with a Post-Optimization phase (SEMPO) to solve it. Extensive computational experiments are conducted on a set of 80 new benchmark instances with up to 183 customers and a challenging time horizon of up 7 to 28 time periods to evaluate the performance of our approaches. The proposed SEMPO algorithm provides high-quality solutions and faster convergence compared to the MILP formulation.}, archive = {J_EJOR}, author = {Diego Perdigão Martino and Philippe Lacomme and Katyanne Farias}, doi = {10.1016/j.ejor.2025.05.044}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {91-104}, shortjournal = {Eur. J. Oper. Res.}, title = {A split-embedded metaheuristic for the heterogeneous inventory routing problem with batch size}, volume = {328}, year = {2026}, }
• (2026). The role of scarcity behavior in inventory management. EJOR, 328(1), 78-90. (www)

  @article{ , abstract = {A well-known phenomenon related to inventory, but often neglected in inventory management, is the scarcity effect, i.e., the increase in the demand if inventory is low. We consider a repeated purchase setting with a single firm and multiple buyers and address the question of whether inventory management decisions concerning the control policy and the service configuration (determining when and how much to order) impact scarcity behavior arising from buyers’ stock-out perception. We study common inventory control policies that assume a demand independent of the inventory management, and we challenge this critical assumption of an exogenous demand. This research explores two prevalent classes of inventory policies widely used in practice (periodic and continuous) configured according to fill rates, a popular way of measuring service. We conduct a laboratory experiment with four automated inventory management treatments (2 policies × 2 configurations) where participants act as buyers. We observe stock-out induced scarcity and find support for the hypothesis that the periodic policy leads to a stronger effect compared to the continuous policy if the service level is low. The study also supports the hypothesis that buyers act forward-looking as their demand peaks before the inventory reaches its lowest level. Overall, our research provides a new perspective on inventory management as it reveals that the chosen control policy and the selected service configuration influence stock-out pressure induced inventory runs. Inventory managers should be aware of scarcity effects and its consequences, like the disadvantages of a periodic policy for low service level.}, archive = {J_EJOR}, author = {Sebastian Schiffels and Christian Jost}, doi = {10.1016/j.ejor.2025.05.043}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {78-90}, shortjournal = {Eur. J. Oper. Res.}, title = {The role of scarcity behavior in inventory management}, volume = {328}, year = {2026}, }
• (2026). Green scheduling with time-of-use tariffs and machine states: Optimizing energy cost via branch-and-bound and bin packing strategies. EJOR, 328(1), 64-77. (www)

  @article{ , abstract = {This paper presents a branch-and-bound algorithm, enhanced with bin packing strategies, for scheduling under variable energy pricing and power-saving states. The proposed algorithm addresses the 1 , TOU | states | TEC problem, which involves scheduling jobs to minimize total energy cost (TEC) while considering time-of-use (TOU) electricity prices and different machine states (e.g., processing, idle, off). Key innovations include instance pre-processing for rapid lower bound calculations, a novel branching scheme combined with initializations, a block-finding primal heuristic, and a tighter lower bound for jobs with non-coprime processing times. These enhancements result in an efficient algorithm capable of solving benchmark instances with real energy prices with 200 jobs more than 100 times faster than existing state-of-the-art methods.}, archive = {J_EJOR}, author = {Ondřej Benedikt and István Módos and Antonin Novak and Zdeněk Hanzálek}, doi = {10.1016/j.ejor.2025.06.026}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {64-77}, shortjournal = {Eur. J. Oper. Res.}, title = {Green scheduling with time-of-use tariffs and machine states: Optimizing energy cost via branch-and-bound and bin packing strategies}, volume = {328}, year = {2026}, }
• (2026). Flight test scheduling: A generic model, lower bounds, and iterated local search. EJOR, 328(1), 49-63. (www)

  @article{ , abstract = {Flight tests play a critical role in the R&D process for new aircraft as they help verify the airworthiness and capabilities and expose design and manufacturing defects. During the test course, a large number of tasks need to be scheduled appropriately so that the flight tests can be completed with minimum cost and high efficiency. Therefore, there is a strong need for developing an efficient method that can generate high-quality test schedules. In this paper, we study flight test scheduling to minimize the number of required test flights, thereby decreasing the cost and time required during the entire test course. We establish a mixed-integer programming model to formally describe the problem, propose several computationally efficient lower bounds to help verify the quality of obtained solutions, and develop an iterated local search algorithm for generating a high-quality solution in an effective manner. Comprehensive computational experiments are performed to demonstrate the efficiency of our proposed algorithm. We report some general managerial insights based on the obtained computational results.}, archive = {J_EJOR}, author = {Hanqiao Tao and Guopeng Song and Roel Leus and Zhe Liang and Jiang Jiang}, doi = {10.1016/j.ejor.2025.06.001}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {49-63}, shortjournal = {Eur. J. Oper. Res.}, title = {Flight test scheduling: A generic model, lower bounds, and iterated local search}, volume = {328}, year = {2026}, }
• (2026). Nested logic-based benders decomposition for an integrated home healthcare problem. EJOR, 328(1), 32-48. (www)

  @article{ , abstract = {In this study, we apply nested logic-based Benders decomposition to solve an integrated home healthcare staffing, assignment, routing, and scheduling problem with application in Norway. The proposed method operates at two decomposition levels. Consequently, the entire problem is decomposed into three hierarchical sub-problems: the staffing problem, the assignment problem, and the routing and scheduling problem. These sub-problems are interrelated through two levels of logic-based Benders cuts. Computational experiments on 40 test instances demonstrate the superior performance of nested logic-based Benders decomposition compared to directly solving a mixed-integer linear programming model available in the literature. Specifically, the proposed solution method achieved proven optimality in 28 instances and provided feasible solutions for the remaining 12 instances. In contrast, directly solving the mixed-integer linear programming model yielded proven optimality in 16 instances, provided feasible solutions for 20 instances, and failed to find feasible solutions for 4 instances within the same computational time limit.}, archive = {J_EJOR}, author = {Abdalrahman Algendi and Sebastián Urrutia and Lars Magnus Hvattum and Rafael A. Melo}, doi = {10.1016/j.ejor.2025.06.006}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {32-48}, shortjournal = {Eur. J. Oper. Res.}, title = {Nested logic-based benders decomposition for an integrated home healthcare problem}, volume = {328}, year = {2026}, }
• (2026). A reduction approach for the parallel machine scheduling problem with a separate server for loading and unloading operations. EJOR, 328(1), 15-31. (www)

  @article{ , abstract = {This paper addresses the non-preemptive identical parallel machine scheduling problem with a dedicated loading server and a dedicated unloading server. Each job has to be loaded by the dedicated loading server immediately before being processed on one of the identical parallel machines, and unloaded immediately by the dedicated unloading server after its processing. The objective function involves the minimization of the makespan. This problem arises in the semiconductor industry, plastic injection industry, kitchen production systems, healthcare, container terminals, and many other industrial fields. We prove the problem to be strongly NP-hard, analyze a special case with identical loading, processing, and unloading times, and establish a tight lower bound. In addition, we propose two novel mixed-integer linear programming formulations: one utilizing time-indexed variables with an iterative strengthening algorithm, and the other employing linear-ordering variables along with two enhanced valid inequalities. Given the complexity of the problem, we introduce a reduction approach that simplifies the problem by modifying certain constraints, enabling the determination of a feasible solution much more quickly. Building on this reduction, we provide a linear-time reduction algorithm and a fast formulation based on assignment-and-positional date variables. Furthermore, we study a special case involving regular jobs. To solve large-scale instances of the problem with up to 250 jobs and up to 5 machines, we design a hybrid approach combining an iterated greedy algorithm with variable neighborhood descent. As shown in the computational experiments on two sets of benchmark instances, the reduction approach significantly outperforms all previous methods existing in the literature.}, archive = {J_EJOR}, author = {Abdelhak Elidrissi and Jatinder N.D. Gupta and Rachid Benmansour and Bertrand M.T. Lin}, doi = {10.1016/j.ejor.2025.05.030}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {15-31}, shortjournal = {Eur. J. Oper. Res.}, title = {A reduction approach for the parallel machine scheduling problem with a separate server for loading and unloading operations}, volume = {328}, year = {2026}, }
• (2026). Cutting stock problem with usable leftovers: A review. EJOR, 328(1), 1-14. (www)

  @article{ , abstract = {This article presents a comprehensive literature review of the Cutting Stock Problem with Usable Leftovers (CSPUL). The most recent review on this topic dates to 2014, covering articles published before 2013. Since then, the number of publications on CSPUL has increased significantly, driven by new applications and more efficient solution approaches. We analyze fifty two relevant articles from twenty four different journals, focusing on works published after 2008 while acknowledging foundational contributions from the 1980s and 1990s. This review categorizes variations of CSPUL based on their dimensions (1D, 2D, and 3D), planning period characteristics (single-period and multi-period), objective functions, and solution methods. The article provides a detailed summary of the key features in the mathematical models and solution methods proposed in these studies. Additionally, it highlights several industrial applications of CSPUL, illustrating its practical relevance. Through this analysis, we identify important applications and propose promising directions for future research. The findings and insights presented here have practical implications for optimizing resource utilization and promoting sustainability in industries facing cutting challenges.}, archive = {J_EJOR}, author = {Victor Senergues and Nadjib Brahimi and Adriana Cristina Cherri and François Klein and Olivier Péton}, doi = {10.1016/j.ejor.2025.03.014}, journal = {European Journal of Operational Research}, month = {1}, number = {1}, pages = {1-14}, shortjournal = {Eur. J. Oper. Res.}, title = {Cutting stock problem with usable leftovers: A review}, volume = {328}, year = {2026}, }

ESWA - 327

• (2026). GRAG-ProSafe QAS: Graph retrieval-augmented generation for process production safety intelligent question and answer system. ESWA, 298, 129947. (www)

  @article{ , abstract = {To address the frequent occurrence of safety accidents in process industries and the inadequacy of traditional safety management approaches in coping with complex risk scenarios, this study proposed an intelligent question answering system based on graph retrieval-augmented generation, named GRAG-ProSafe QAS. The system adopted a four-stage framework to extract structured knowledge from accident reports using Large Language Models and constructed a dynamic knowledge graph. This graph is utilized to support multi-hop retrieval and chain-of-thought reasoning strategies, thereby enhancing the trustworthiness of the generated responses. Taking 198 accident reports from the iron and steel industry as the dataset, the experiments demonstrated that the model achieved efficient structured knowledge extraction, and the constructed knowledge graph covers 1637 nodes with 2285 edges, which can provide semantic support for multi-hop causal reasoning. Ablation experiments validated the effectiveness of each core module in the system such as structured graph retrieval. In the multi-turn question answering scenario focused on explosion accidents, the system achieved strong overall performance, with a Faithfulness Score of 0.8684, Answer Relevancy of 0.8241, and Factual Correctness of 0.8050. The developed GRAG-ProSafe QAS realized the integration of knowledge extraction, graph construction and intelligent QA, which provides a feasible path for the intelligent management and control of safety risks in the process industry.}, archive = {J_ESWA}, author = {Jianrong Zhang and Wei Zhang and Qingwen Wei and Huayu Zhong and Yu Miao}, doi = {10.1016/j.eswa.2025.129947}, journal = {Expert Systems with Applications}, month = {3}, pages = {129947}, shortjournal = {Expert Syst. Appl.}, title = {GRAG-ProSafe QAS: Graph retrieval-augmented generation for process production safety intelligent question and answer system}, volume = {298}, year = {2026}, }
• (2026). Relevance-based adaptive differential private spiking neural networks. ESWA, 298, 129934. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs) have achieved widespread attention recently due to biological plausibility and ultra-low energy consumption. However, the training process of SNNs can potentially cause data privacy leaks. Attackers can use models to infer sensitive features of data, therefore, protecting sensitive information in training data is crucial. Differential Privacy (DP) is an effective method for protecting data privacy from attacks. The existing differential private SNN models are not satisfactory due to adding the same amount of noise to gradient parameters, which affects the trade-off between model utility and privacy protection. To address these issues, a Relevance-based Adaptive Differential Private SNN (RADP-SNN) framework is proposed, which dynamically perturbs gradients based on the relevance between input features and model outputs. It adjusts the Gaussian noise added to the gradients according to each neuron’s relevance to the model outputs. Additionally, based on the interpretability of the Layer-wise Relevance Propagation (LRP) algorithm, an SNNLRP algorithm is designed, which uses surrogate gradients to approximate the backpropagation of spiking neurons. It integrates the spikes and membrane potential information across time steps in SNNs. Then, the impact of Leaky-Integrated and Fire and Integrated and Fire neuron models on privacy protection is explored. Results show that the proposed framework can maintain high model performance while providing effective privacy assurance. RADP-SNN has broad potential applications in image, robot control, and scenarios requiring time series data processing.}, archive = {J_ESWA}, author = {Junxiu Liu and Xiwen Luo and Qiang Fu and Yuling Luo and Sheng Qin and Xue Ouyang}, doi = {10.1016/j.eswa.2025.129934}, journal = {Expert Systems with Applications}, month = {3}, pages = {129934}, shortjournal = {Expert Syst. Appl.}, title = {Relevance-based adaptive differential private spiking neural networks}, volume = {298}, year = {2026}, }
• (2026). EDPDet: Efficient dense pedestrian detectors with multi-scale feature extraction, enhancement and aggregation. ESWA, 298, 129933. (www)

  @article{ , abstract = {Pedestrian detection is one of the key tasks in the field of computer vision. However, in dense scenes, pedestrians may be confused with the background or obstructed to varying degrees, and factors such as shooting distance can affect the scale of pedestrians in images, increasing the difficulty of pedestrian detection tasks. To address these issues, we propose the efficient dense person detectors called EDPDet. On the one hand, we propose the Multi-scale Contour Feature Enhancement Module (MCFEM) with multi-scale feature extraction and contour feature enhancement, which can improve the network’s ability to perceive pedestrian objects at different scales and distinguish pedestrians from complex backgrounds. On the other hand, we propose the Efficient Cross-scale Path Aggregation Network (ECPAN) that realizes the integration of semantic information and spatial information among different feature layers of the network. Furthermore, we implemented the end-to-end deployment of the detectors using the non-maximum suppression-free (NMS-free) training strategy. Finally, we construct three detectors of different sizes, EDPDet-N, EDPDet-S and EDPDet-L. Our method achieves 91.4% and 92.5% AP on the CrowdHuman and WiderPerson datasets, respectively, demonstrating superior performance.}, archive = {J_ESWA}, author = {Yangxi Yu and Wentao Lyu and Qing Guo and Zhijiang Deng and Weiqiang Xu}, doi = {10.1016/j.eswa.2025.129933}, journal = {Expert Systems with Applications}, month = {3}, pages = {129933}, shortjournal = {Expert Syst. Appl.}, title = {EDPDet: Efficient dense pedestrian detectors with multi-scale feature extraction, enhancement and aggregation}, volume = {298}, year = {2026}, }
• (2026). SketchFormer3D: Generating 3D shapes from sketches with implicit SDF priors via diffusion models. ESWA, 298, 129931. (www)

  @article{ , abstract = {As an intuitive and highly efficient means of conveying information, sketches are widely used in three-dimensional design and modeling tasks. However, due to the highly abstract and visually sparse nature of sketches, directly reconstructing structurally clear and topologically sound 3D models from single-view sketches remains a significant challenge. Existing approaches often overlook the cross-modal semantic associations between sketches and 3D geometry and fail to effectively utilize prior knowledge of three-dimensional spatial geometry, resulting in deficiencies in the structural coherence and topological validity of the generated 3D models. To address these issues, we propose a 3D reconstruction method based on a Signed Distance Function (SDF) diffusion model, aiming to enhance the stability and accuracy of model generation through implicit geometric priors and cross-modal semantic information. Specifically, we first employ an SDF-based implicit neural network to construct high-quality geometric representations and capture spatial priors of target shapes; this component is then incorporated as a fixed module into the diffusion process, guiding the network to gradually recover accurate 3D structures from noise under sketch conditions. Meanwhile, we design a sketch-aware semantic encoding module and a cross-modal attention mechanism to deeply integrate sketch images with geometric priors, thereby strengthening the model’s ability to represent structural consistency and local details. Experimental results demonstrate that our method outperforms existing approaches across multiple sketch-based modeling datasets. In terms of quantitative metrics, compared to the latest method Doodle Your 3D, our Chamfer Distance (CD) decreases from 9.41 × 10 − 3 to 6.53 × 10 − 3 , the Earth Mover’s Distance (EMD) decreases from 10.68 × 10 − 2 to 7.59 × 10 − 2 , and the Voxel-IoU increases from 46.19 × 10 − 2 to 48.72 × 10 − 2 . These improvements further validate that our approach is capable of stably generating 3D mesh models that are both structurally coherent and geometrically precise.}, archive = {J_ESWA}, author = {Yang Ding and Huamin Yang and Cheng Han and Chao Zhang and Yu Liu}, doi = {10.1016/j.eswa.2025.129931}, journal = {Expert Systems with Applications}, month = {3}, pages = {129931}, shortjournal = {Expert Syst. Appl.}, title = {SketchFormer3D: Generating 3D shapes from sketches with implicit SDF priors via diffusion models}, volume = {298}, year = {2026}, }
• (2026). Modeling trust in human–Robot collaborative construction: An improved cloud bayesian network. ESWA, 298, 129928. (www)

  @article{ , abstract = {Trust developed by workers towards robotic systems is critical to the successful implementation of human-robot collaboration (HRC) in construction, directly influencing operational efficiency and safety outcomes. To accurately evaluate trust risks within HRC scenarios, this study proposes an integrated method combining an improved Cloud Model (CM) with Bayesian Networks (BNs) for dynamic trust risk analysis. Initially, key factors influencing trust risks in HRC were identified through literature review and expert elicitation. The improved CM was then employed to capture inherent uncertainties and fuzziness in trust state definitions, facilitating the discretization of continuous expert evaluations into appropriate risk states. Subsequently, the BN was developed to perform forward reasoning, sensitivity analysis, and backward diagnosis, enabling proactive trust risk prediction, critical factor identification, and targeted interventions. The primary contributions of this research include: (a) identifying 11 trust factors from human, organizational, and robotic perspectives, offering a comprehensive basis for analyzing HRC trust risk in construction; (b) employing an optimized cloud entropy approach to accurately capture fuzziness and randomness in expert evaluations, thereby producing robust prior probabilities; and (c) developing a hybrid CBN framework to assess HRC trust risk in construction, demonstrating superior performance in risk perception, analysis, and control. Overall, this study provides valuable insights into safer and more effective HRC through dynamic evaluation of trust risk.}, archive = {J_ESWA}, author = {Lei Wang and Mingyu Zhang and Heng Li and Yinong Hu and Jie Ma and Waleed Umer and Xin Fang}, doi = {10.1016/j.eswa.2025.129928}, journal = {Expert Systems with Applications}, month = {3}, pages = {129928}, shortjournal = {Expert Syst. Appl.}, title = {Modeling trust in human–Robot collaborative construction: An improved cloud bayesian network}, volume = {298}, year = {2026}, }
• (2026). Automated emotional design generation for NEV wheel hubs: Integrating StyleGAN2-ADA and WOA-SVR within kansei engineering. ESWA, 298, 129927. (www)

  @article{ , abstract = {Wheel hub design is significant in the New Energy Vehicle’s (NEV) aesthetics and user emotional connection. However, the existing design process faces two main challenges: traditional concept methods suffer from efficiency bottlenecks, and designers’ subjective judgments make it difficult to capture users’ emotional preferences. Recently, Generative Adversarial Networks (GANs) have been introduced to industrial design. However, conventional GAN-based methods typically require large-scale datasets to achieve high-quality design generation, which limits their applicability in vertical product domains where data are scarce. To overcome these challenges, this paper proposes an emotional design generation approach for NEV wheel hubs by combining StyleGAN2 with Adaptive Discriminator Augmentation (StyleGAN2-ADA) and Whale Optimisation Algorithm-Support Vector Regression (WOA-SVR). First, the StyleGAN2-ADA is trained on a limited NEV wheel hub dataset to achieve automated design generation. Second, Ward hierarchical clustering dentifies representative samples, while morphological analysis deconstructs design features. Subsequently, Factor Analysis (FA) categorises Kansei words to extract principal emotions. After collecting users’ emotional ratings of wheel hubs, WOA-SVR constructs a “design features-emotional needs” mapping model. Finally, the generated images are used to validate the implementation of the generation and prediction models. The case study demonstrates that the proposed method not only generates diverse design alternatives aligned with users’ emotional preferences but also provides reliable predictions of their emotional responses, thereby systematically improving the NEV wheel hub design process while preserving its emotional impact.}, archive = {J_ESWA}, author = {Yi Wang and Meiyu Zhou and Zhengyu Wang and Weilin Cai and Xin Sun and Huijuan Zhu}, doi = {10.1016/j.eswa.2025.129927}, journal = {Expert Systems with Applications}, month = {3}, pages = {129927}, shortjournal = {Expert Syst. Appl.}, title = {Automated emotional design generation for NEV wheel hubs: Integrating StyleGAN2-ADA and WOA-SVR within kansei engineering}, volume = {298}, year = {2026}, }
• (2026). DiffAD: Diffusion augmentation for influenza virus antigenic distance prediction. ESWA, 298, 129925. (www)

  @article{ , abstract = {Continuous monitoring of the antigenic phenotypes of circulating influenza viruses is essential for both vaccine evaluation and development. However, traditional serology-based methods for antigenicity assessment, such as hemagglutination inhibition assays, are inherently low-throughput and labor-intensive. As a result, the available antigenicity data remain sparse and limited, posing a significant challenge to the development of accurate and scalable computational models. To address this, we have curated a influenza antigenicity dataset from authoritative sources and propose DiffAD, a novel diffusion-based data augmentation framework for antigenic distance prediction to enhance data availability and model performance. DiffAD introduces a forward noising process and a timestep-aware reverse denoising network incorporating pre-normalized residual blocks to enable precise noise estimation across diffusion steps. This design facilitates the generation of biologically plausible HA feature vectors that introduce meaningful variability and augment existing datasets. These augmented data are then used to enhance the accuracy of antigenic distance estimation between viral strains. Experimental results on four influenza subtypes (H1N1, H3N2, B/Victoria, and B/Yamagata), using two widely adopted antigenic distance metrics. Experimental results show that DiffAD significantly outperforms existing methods, reducing the MAE to 0.6402 and 0.6795, and increasing the R 2 to 0.7412 and 0.6969 on the H1N1 and H3N2 datasets, respectively, using the SLT-based distance metric. Furthermore, visualization analysis reveals that the low-dimensional embeddings of virus strains exhibit clear clustering patterns, and the synthetic strains generated by DiffAD closely align with original antigenic clusters. These findings demonstrate that DiffAD not only improves prediction performance but also learns biologically meaningful characteristics of viral strains.}, archive = {J_ESWA}, author = {Qingbo Liu and Yuanling Xia and Yanbu Guo and Weihua Li}, doi = {10.1016/j.eswa.2025.129925}, journal = {Expert Systems with Applications}, month = {3}, pages = {129925}, shortjournal = {Expert Syst. Appl.}, title = {DiffAD: Diffusion augmentation for influenza virus antigenic distance prediction}, volume = {298}, year = {2026}, }
• (2026). Learning to select cutting planes in mixed integer linear programming solving. ESWA, 298, 129924. (www)

  @article{ , abstract = {Cutting planes (cuts) are crucial for solving Mixed Integer Linear Programming (MILP) problems. Advanced MILP solvers typically rely on manually designed heuristic algorithms for cut selection, which require extensive expert knowledge and cannot be generalized to MILP problems with varying instance sizes across different domains. Therefore, learning-based methods for cut selection are considered a promising direction. State-of-the-art learning-based methods formulate cut selection as a sequence-to-sequence problem, easily handled by sequence models. However, the existing sequence models need help with the following issues: (1) the model only captures cut information while neglecting the Linear Programming (LP) relaxation; (2) the sequence model utilizes positional information of the input sequence, which may influence cut selection. To address these challenges, we design a novel model HGTSM for better cut selection. We encode MILP problem state as a heterogeneous tripartite graph, utilizing heterogeneous graph networks to fully capture the underlying structure of MILP problems. Simultaneously, we propose a novel sequence model whose architecture is tailored to handle inputs in different orders. Experimental results demonstrate that our model outperforms heuristic methods and learning-based baselines on multiple challenging MILP datasets. The code of HGTSM can be found at https://github.com/aqi007/Learn2Cut-HGTSM.}, archive = {J_ESWA}, author = {Xuefeng Zhang and Liangyu Chen and Zhengfeng Yang and Zhenbing Zeng}, doi = {10.1016/j.eswa.2025.129924}, journal = {Expert Systems with Applications}, month = {3}, pages = {129924}, shortjournal = {Expert Syst. Appl.}, title = {Learning to select cutting planes in mixed integer linear programming solving}, volume = {298}, year = {2026}, }
• (2026). A machine learning-based medical device recall initiator prediction framework: From supply chain risk management and resilience view. ESWA, 298, 129922. (www)

  @article{ , abstract = {Persistent quality problems with medical devices and the associated recall present potential health risks to users, bringing extra costs and disturbances to the supply chain. Classical medical device recall strategy neglects the significance of the failure detection process in the premarket phase, increasing the medical device recall risks. This research first established the theoretical foundation for the medical device recall reasons detection problem by reconstructing the medical device recall strategy from the supply chain risk and resilience view and reinforced the importance of failure detection and quality inspection work in the premarket stage. Moreover, existing medical device failure reason prediction research was limited in practicality and scalability. To address this problem, we developed a machine learning-based medical device recall initiator prediction system framework to conduct proactive failure detection based on the industrial case. By redesigning in dataset, clustering method and input feature selection, an accuracy rate of 88.85% is achieved, which indicates the potential of the proposed framework in assisting manufacturers with asset predictive failure detection for reducing recall. A comparative analysis of prediction performance between our framework and the most similar research that utilized the same prediction algorithms was presented. The comparison results showed that our distinctive design in the dataset, clustering method, and key input features chosen are valid and efficient. Before redesigning the prediction algorithms that require higher technical investment, our elaborate research design in selecting the dataset, cluster method, and key input features can be the antecedents of better prediction performance for manufacturers. The proposed predictive framework obtains higher accuracy, scalability, practicality, with accessibility.}, archive = {J_ESWA}, author = {Yang Hu and Davy Monticolo and Pezhman Ghadimi}, doi = {10.1016/j.eswa.2025.129922}, journal = {Expert Systems with Applications}, month = {3}, pages = {129922}, shortjournal = {Expert Syst. Appl.}, title = {A machine learning-based medical device recall initiator prediction framework: From supply chain risk management and resilience view}, volume = {298}, year = {2026}, }
• (2026). Predicting last-mile delivery route deviations using machine learning. ESWA, 298, 129921. (www)

  @article{ , abstract = {Route planning in last-mile delivery is a complex task with many challenges, directly impacting delivery efficiency and costs. Drivers often deviate from optimized planned routes based on their knowledge. Using the properties of machine learning, this study aims to determine whether machine learning techniques can effectively predict deviations by drivers from planned routes and quantify the extent of such deviations. We propose to predict route deviations by analyzing a logistics company’s historical data of planned and actual routes using deep neural networks, with the dataset made publicly available. Our methodology incorporates both regression and classification models. The regression model estimates the degree of deviation, while the classification model aims to predict whether the deviation from a planned route will exceed a given threshold, based on different deviation metrics. As the input, we leverage the sequential structure of the route with route properties and drivers information. The computational experiments explore extending the given input to the models and testing various state-of-art neural network architectures. Our results demonstrate strong performance on both tasks, with our models achieving 9 − 19 % improvements in regression metrics and 3 − 15 % improvements in classification metrics compared to specified benchmarks, with statistical tests confirming the significance of these improvements.}, archive = {J_ESWA}, author = {Anna Konovalenko and Lars Magnus Hvattum and Kim Aleksander Hammer Iversen}, doi = {10.1016/j.eswa.2025.129921}, journal = {Expert Systems with Applications}, month = {3}, pages = {129921}, shortjournal = {Expert Syst. Appl.}, title = {Predicting last-mile delivery route deviations using machine learning}, volume = {298}, year = {2026}, }
• (2026). PDHG: An ethereum phishing detection approach via heterogeneous graph transformer. ESWA, 298, 129919. (www)

  @article{ , abstract = {Phishing scams have emerged as a significant threat within the Ethereum ecosystem. Cutting-edge Ethereum phishing scams detection techniques mostly treat accounts in Ethereum as homogeneous nodes in transaction graphs. Existing detection approaches model Ethereum transaction records as homogeneous transaction graphs and use graph representation learning for account classification. However, those approaches often overlook the heterogeneity between accounts and transactions, making it difficult to capture the diversity of interactions and features among accounts. In this paper, a heterogeneous graph transformer (HGT)-based phishing account identification approach called PDHG is proposed. Specifically, PDHG models the transaction network between accounts as a heterogeneous graph based on different attributes of Ethereum accounts, allowing for a more comprehensive description of the structure and behavioral patterns of the transaction network. To enhance the explainability, PDHG leverages PDHGexplainer as the explainer for the detection results. We compare PDHG with other existing detection models. The experimental results demonstrate that PDHG achieves an AUC score of 96.04 % and a recall score of 89.87 %, surpassing the state-of-the-art approaches.}, archive = {J_ESWA}, author = {Lei Wang and Yihan Mi and Yanan Zhang and Jialin Zhang}, doi = {10.1016/j.eswa.2025.129919}, journal = {Expert Systems with Applications}, month = {3}, pages = {129919}, shortjournal = {Expert Syst. Appl.}, title = {PDHG: An ethereum phishing detection approach via heterogeneous graph transformer}, volume = {298}, year = {2026}, }
• (2026). Multi-graph fusion guided robust adaptive learning for subspace clustering. ESWA, 298, 129918. (www)

  @article{ , abstract = {Subspace clustering is an advanced technique that identifies clusters embedded within a union of low-dimensional subspaces of the original data space, thereby revealing its intrinsic structure. Spectral clustering-based methods have gained significant attention in computer vision, image processing and pattern recognition due to their promising performance. However, existing approaches, which typically rely on self-representation for representation coefficient learning, often lack robustness and struggle to comprehensively characterize complex data structures. Traditional reconstruction loss based on the Frobenius or ℓ 1 norm are susceptible to noise and outliers. Furthermore, many methods underutilize inherent data characteristics for capturing local geometric structures and adapting to intricate data relationships. To address these limitations, this paper proposes a novel subspace clustering approach, named Multi-graph Fusion Guided Robust Adaptive Learning (MFGRAL), which integrates robust adaptive representation and multi-graph fusion within a unified framework. Specifically, a non-convex logarithmic loss function is adopted to enhance robustness against noise and outliers. To better preserve local manifold structures, a multi-graph fusion strategy is developed to guide the adaptive graph learning process. This facilitates the learning of more discriminative low-dimensional embeddings and enhances the capacity to capture complex neighborhood relationships. An effective and efficient optimization algorithm based on Alternating Direction Method of Multipliers (ADMM) is developed to solve the proposed model. Extensive experimental results on several benchmark datasets demonstrate the effectiveness of the proposed MFGRAL and its superiority over state-of-the-art methods.}, archive = {J_ESWA}, author = {Jianyu Miao and Xiaochan Zhang and Chao Fan and Tiejun Yang and Yingjie Tian and Yong Shi and Mingliang Xu}, doi = {10.1016/j.eswa.2025.129918}, journal = {Expert Systems with Applications}, month = {3}, pages = {129918}, shortjournal = {Expert Syst. Appl.}, title = {Multi-graph fusion guided robust adaptive learning for subspace clustering}, volume = {298}, year = {2026}, }
• (2026). Enhancing storage efficiency for cloud-based vehicle data based on advanced lossless compression algorithm. ESWA, 298, 129917. (www)

  @article{ , abstract = {Cloud-based vehicle data is being produced in various forms at an unparalleled speed, which requires efficient compression mechanisms to better store, transmit and process such data. In this article, we devise a novel, hybrid, adaptation-enabled framework, RTLAE, built on a neural network (NN) predictor and an adaptive arithmetic coding (AC) encoder, which enables outstanding lossless compression performance on non-stationary (unanticipated contextual changes) or randomly mixed vehicle datasets without requiring any domain-specific prior knowledge (lacking optimal hyperparameter settings). This framework integrates a tailor-made redundancy reduction filter, a temporal convolutional network, a long short-term memory network, and a multi-order adaptive arithmetic encoder to preprocess data sequences, generate accurate estimations, and learn optimal character encoding. Validation results on real-life datasets indicate that the proposed framework has achieved quite superior results on various datasets, with an average performance improvement of about 26%, 46%, 27%, and 12% over AC, Gzip, block-sorting compression algorithm, and DeepZip, respectively, and with a runtime cost that is slightly better than other NN-based compressors. This work emphasizes the potential of NNs and combinatorial modeling to improve general-purpose compressors, and highlights the research prospect of resource utilization of big data platforms based on vehicle-cloud collaborative interconnection.}, archive = {J_ESWA}, author = {Zheng Yifan and Cao Rui and Liu Chaohui and Liu Qixing and Luo Zixuan and Yang Dong and Zhang Kun and Zhou Sida and Yang Shichun}, doi = {10.1016/j.eswa.2025.129917}, journal = {Expert Systems with Applications}, month = {3}, pages = {129917}, shortjournal = {Expert Syst. Appl.}, title = {Enhancing storage efficiency for cloud-based vehicle data based on advanced lossless compression algorithm}, volume = {298}, year = {2026}, }
• (2026). Complex forecasting and investment strategy optimization via chain-of-thought of large language models. ESWA, 298, 129913. (www)

  @article{ , abstract = {This paper pioneers the application of chain-of-thought (CoT) prompting in large language models (LLMs) for financial forecasting and portfolio optimization. Leveraging anonymized financial statements from 608 Chinese A-share listed companies (2010–2023), we benchmark ChatGPT 4.0 against human analysts in earnings direction prediction. The CoT-enhanced model achieved superior accuracy (64.35% vs. 58.37%) and generated a 17.14% alpha with a Sharpe ratio of 1.5959 in backtests, demonstrating LLMs’ capacity to automate financial reasoning while reducing reliance on specialized expertise. Our findings bridge AI and quantitative finance by validating LLMs’ cross-domain adaptability using purely numerical data, offering practical implications for AI-driven investment decision systems.}, archive = {J_ESWA}, author = {Meiqun Yin and Mengzhu Guo}, doi = {10.1016/j.eswa.2025.129913}, journal = {Expert Systems with Applications}, month = {3}, pages = {129913}, shortjournal = {Expert Syst. Appl.}, title = {Complex forecasting and investment strategy optimization via chain-of-thought of large language models}, volume = {298}, year = {2026}, }
• (2026). KMHBO: A knowledge-guided multi-niche hybrid breeding optimization algorithm for high-dimensional multimodal feature selection. ESWA, 298, 129911. (www)

  @article{ , abstract = {Multimodal feature selection aims to identify multiple feature subsets with similar discriminative capabilities, providing diverse and interchangeable solutions for intelligent decision-making. However, most existing multimodal feature selection methods based on metaheuristic algorithms are designed for low or medium dimensional data, and their search efficiency significantly degrades in high-dimensional spaces. To address this challenge, we propose a knowledge-guided multi-niche hybrid breeding optimization algorithm (KMHBO) with dual dynamic grouping, which enhances both solution diversity and overall search performance. KMHBO introduces a two-level dynamic grouping mechanism: at the outer level, a dual similarity-guided niche partitioning strategy adaptively adjusts the size and distribution of each niche to improve the global exploration; at the inner level, the population is divided into three subpopulations based on historical evolutionary behaviors, namely the maintainer, the restorer, and the sterile. Each subpopulation is assigned specific search tasks to strengthen the local exploitation and increase structural diversity among solutions. Additionally, a mutation mechanism based on feature selection frequency is introduced to promote information exchange between niches, while an improved external archive updating mechanism ensures the preservation of diverse equivalent solutions. Extensive experiments on ten high-dimensional benchmark datasets demonstrate that KMHBO significantly outperforms state-of-the-art methods in terms of classification accuracy. Notably, KMHBO achieves 100% classification accuracy on the ALLAML and GLI_85 datasets and identifies nearly 50 structurally distinct yet functionally equivalent feature subsets under a fixed archive size, showcasing its superior performance and strong potential in high-dimensional multimodal feature selection tasks.}, archive = {J_ESWA}, author = {Zhiwei Ye and Ying Zeng and Ting Cai and Jun Shen and Wen Zhou and Qiyi He and Mengqing Mei and Yi Cen}, doi = {10.1016/j.eswa.2025.129911}, journal = {Expert Systems with Applications}, month = {3}, pages = {129911}, shortjournal = {Expert Syst. Appl.}, title = {KMHBO: A knowledge-guided multi-niche hybrid breeding optimization algorithm for high-dimensional multimodal feature selection}, volume = {298}, year = {2026}, }
• (2026). A fault-tolerant task offloading framework via large-scale multi-objective evolutionary optimization and game-based decision mechanism. ESWA, 298, 129910. (www)

  @article{ , abstract = {Large-scale multi-access edge computing (MEC) optimization is challenging due to high-dimensional decision spaces, conflicting objectives, nonstationary conditions, and failure-prone infrastructure. This paper presents an adaptive Mahalanobis distance-based large-scale multi-objective evolutionary algorithm with knowledge transfer and a two-layer encoding ( ALC-LSMOEA-KT ). The task-offloading model optimizes latency, energy, load balance, and failure risk under communication and computation constraints. A two-layer sparse encoding separates variable activation from value search, and a phase-aware evolution with Mahalanobis-guided covariance adaptation exploits inter-variable correlations while preserving diversity. A Stackelberg-based fault-tolerant migration module reassigns disrupted tasks to sustain robustness. Experiments on scalable multi-objective optimization Problems(SMOP)/large-scale multi-objective optimization problem(LSMOP) benchmarks and a realistic MEC simulator with dynamic arrivals, bandwidth variation, and injected failures show consistent gains in inverted generational distance (IGD), solution diversity, and robustness. The results indicate a scalable and reliable approach to MEC optimization under high dimensionality and uncertainty.}, archive = {J_ESWA}, author = {Tingting Dong and Jinbu Wen and Fei Xue and Yuge Geng and Xingjuan Cai}, doi = {10.1016/j.eswa.2025.129910}, journal = {Expert Systems with Applications}, month = {3}, pages = {129910}, shortjournal = {Expert Syst. Appl.}, title = {A fault-tolerant task offloading framework via large-scale multi-objective evolutionary optimization and game-based decision mechanism}, volume = {298}, year = {2026}, }
• (2026). A computational learning pipeline for glaucoma progression detection based on the prediction of visual field changes from fundus photographs. ESWA, 298, 129907. (www)

  @article{ , abstract = {Detection of glaucoma progression is crucial to managing patients, permitting individualized care plans and treatment. It is a challenging task requiring the assessment of structural changes to the optic nerve head and functional changes based on visual field testing. Artificial intelligence, especially deep learning techniques, has shown promising results in many applications, including glaucoma diagnosis. This paper proposes a two-stage computational learning pipeline for detecting glaucoma progression using only fundus photographs. In the first stage, a deep learning model takes a time series of fundus photographs as input and outputs a vector of predictions where each element represents the overall rate of change in visual field (VF) sensitivity values for a sector (region) of the optic nerve head (ONH). We implemented two deep learning models—ResNet50 and InceptionResNetV2—for this stage. In the second stage, a binary classifier (weighted logistic regression) takes the predicted vector as input to detect progression. We also propose a novel method for constructing annotated datasets from temporal sequences of clinical fundus photographs and corresponding VF data suitable for machine learning. Each dataset element comprises a temporal sequence of photographs together with a vector-valued label. The label is derived by computing the pointwise linear regression of VF sensitivity values at each VF test location, mapping these locations to eight ONH sectors, and assigning the overall rate of change in each sector to one of the elements of the vector. We used a retrospective clinical dataset with 82 patients collected at multiple timepoints over five years in our experiments. The InceptionResNetV2-based implementation yielded the best performance, achieving detection accuracies of 97.28 ± 1.10 % for unseen test data (i.e., each dataset element is unseen but originates from the same set of patients appearing in the training dataset), and 87.50 ± 0.70 % for test data from unseen patients (training and testing patients are entirely different). The testing throughput was 11.60 ms per patient. These results demonstrate the efficacy of the proposed method for detecting glaucoma progression from fundus photographs.}, archive = {J_ESWA}, author = {Md.Reduanul Haque and Andrew Mehnert and William Huxley Morgan and Graham Mann and Ferdous Sohel}, doi = {10.1016/j.eswa.2025.129907}, journal = {Expert Systems with Applications}, month = {3}, pages = {129907}, shortjournal = {Expert Syst. Appl.}, title = {A computational learning pipeline for glaucoma progression detection based on the prediction of visual field changes from fundus photographs}, volume = {298}, year = {2026}, }
• (2026). Symmetric difference consistent semi-supervised brain stroke lesion segmentation with uncertainty guidance. ESWA, 298, 129900. (www)

  @article{ , abstract = {The sudden onset of stroke is a leading cause of death and disability worldwide. Automatic segmentation of stroke lesions is crucial for the timely formulation of clinical treatment strategies. However, the extensive variations in size, shape, and intensity of stroke lesions present significant challenges for precise segmentation, particularly in the absence of a large set of meticulously annotated data for model training. In this paper, we propose a symmetric difference consistent semi-supervised learning method (SDC-SSL) to leverage readily available unlabeled data for stroke lesion segmentation. We first design a symmetric difference network (SD-Net) that incorporates anatomical symmetry priors, with two key modules: difference awareness extraction (DAE) and symmetric feature fusion (SFF), to accurately segment diverse lesions. Specifically, DAE extracts more discriminative features by exploring multi-scale bilateral differences, while SFF adaptively integrates symmetric features enriched with salient pathological information. To excavate meaningful representations from unlabeled data, SD-Net is then embedded into a novel dual-level consistency mean teacher framework, where uncertainty estimation progressively guides the student in acquiring reliable knowledge from the teacher. In addition to pixel-level consistency regularization, we introduce a difference-aware attention consistency loss at feature level to capture robust semantic information. Finally, we specifically develop pseudo-labels driven symmetric contrastive learning to amplify bilateral differences caused by stroke lesions, further enhancing semi-supervised segmentation. Experimental results show that our SDC-SSL consistently outperforms state-of-the-art semi-supervised methods and achieves performance comparable to fully-supervised approaches with only 20% of labeled data, highlighting its potential clinical value in reducing reliance on extensive data annotation.}, archive = {J_ESWA}, author = {Yongliang Zhang and Wei Wang and Lei Li and Qi Wu and Haoquan Sun and Ao Li and Minghui Wang and Feng Gao}, doi = {10.1016/j.eswa.2025.129900}, journal = {Expert Systems with Applications}, month = {3}, pages = {129900}, shortjournal = {Expert Syst. Appl.}, title = {Symmetric difference consistent semi-supervised brain stroke lesion segmentation with uncertainty guidance}, volume = {298}, year = {2026}, }
• (2026). Operator learning-based springback behavior prediction for complex-shaped tube free-bending forming. ESWA, 298, 129899. (www)

  @article{ , abstract = {Free-bending (FB) technology enables the efficient processing of spatially complex-shaped tubes. Springback causes variations in curvature and torsion of the tube axis during the FB process. The mapping relationship of bent tube curvature and torsion from ideal to actual values can be abstracted as nonlinear physical operators. This paper first proposes a novel six-axis FB processing method that can control geometric features of tube transition segments. Then, an operator learning-based springback behavior prediction (OL-SBP) framework is presented, which includes an OL module and an SBP module. A feature-information-enhanced deep operator network (FIE-DeepONet) is integrated into the first module to learn tube springback operators. The curvature and torsion predicted by the OL module are then fed into the SBP module to calculate the overall shape of the springback axis. This paper also introduces a set of similarity evaluation indicators that are independent of the curve’s spatial attitude. Planar and spatial bent tubes are selected as case studies. Results show that the framework yields more accurate predictions compared to the analytical model. The framework also exhibits excellent generalization performance. Once FIE-DeepONet has learned the springback operators, it can accurately predict the springback curvature and torsion, even for tube shapes not present during training.}, archive = {J_ESWA}, author = {Yongzhe Xiang and Zili Wang and Shuyou Zhang and Le Wang and Caicheng Wang and Yaochen Lin and Jianrong Tan}, doi = {10.1016/j.eswa.2025.129899}, journal = {Expert Systems with Applications}, month = {3}, pages = {129899}, shortjournal = {Expert Syst. Appl.}, title = {Operator learning-based springback behavior prediction for complex-shaped tube free-bending forming}, volume = {298}, year = {2026}, }
• (2026). A multi-label chest X-ray image classification agorithm based on multi-scale and attribute-aware semantic graph. ESWA, 298, 129898. (www)

  @article{ , abstract = {Multi-label Chest X-Ray classification is crucial for intelligent diagnosis, yet existing algorithms usually ignore lesion-scale heterogeneity and attribute-conditioned label dependencies, limiting their clinical generalizability. To address these issues, this paper proposes MSASG, a multi-label Chest X-Ray image classification algorithm based on Multi-Scale and Attribute-aware Semantic Graph which enhances discriminative power and semantic consistency. Firstly, a Multi-scale Feature Partitioning and Reconstruction method is proposed to capture lesion patterns at different scales. Secondly, a Label-guided Multi-scale Semantic Alignment method is proposed to improve visual-semantic alignment by integrating label embeddings into feature extraction and using a Transformer to model high-order cross-modal dependencies. Finally, an Attribute-aware Graph Convolutional Network method is proposed to construct attribute-specific label co-occurrence matrices and dynamically select relevant structures during inference, enabling personalized characterization of label dependencies. Experiments on ChestX-ray14 and CheXpert show that MSASG outperforms state-of-the-art methods in recognizing complex lesion co-occurrence and adapting to heterogeneous populations.}, archive = {J_ESWA}, author = {Qian Wang and Zhijuan Wu and Jiyu Gao and Hongnian Yu and Yongqiang Cheng}, doi = {10.1016/j.eswa.2025.129898}, journal = {Expert Systems with Applications}, month = {3}, pages = {129898}, shortjournal = {Expert Syst. Appl.}, title = {A multi-label chest X-ray image classification agorithm based on multi-scale and attribute-aware semantic graph}, volume = {298}, year = {2026}, }
• (2026). ABE-mamba: Few-shot medical image segmentation via adversarial bidirectional enhanced mamba. ESWA, 298, 129897. (www)

  @article{ , abstract = {Traditional Few-Shot Medical Image Segmentation (FSMIS) methods primarily focus on mining information from support image to guide query image segmentation, while insufficiently exploring the contextual information within the query image itself. Due to significant anatomical variability and differences in imaging conditions, the targets in support and query images may exhibit substantial discrepancies, even within the same class. Consequently, inadequate utilization of query context limits the model’s ability to adapt to intra-class variations, reducing its generalization capability. To address this issue, inspired by Mamba’s capacity for context modeling with linear complexity, we propose ABE-Mamba, a novel query-focused FSMIS framework. ABE-Mamba enhances the model’s adaptability to intra-class variations by effectively mining the query image’s contextual information, and introduces an adversarial training mechanism to improve high-order consistency between the predicted mask and the ground truth. Firstly, a Bidirectional Enhanced Mamba module is introduced into the generator to fully mine the rich semantics in query images by facilitating bidirectional enhancement of both local and global context. Secondly, a cross 2D Selective Scan (cross-SS2D) block is designed and incorporated into the discriminator to improve its capability in distinguishing subtle differences between real and generated masks, thereby enhancing the alignment between predicted masks and ground truths. Thirdly, a pyramid structure is integrated into the generator to facilitate multi-scale feature extraction, thereby improving robustness to objects of varying sizes. Extensive experiments on four widely used datasets (Syn-CT, CHAOS-MRI, Card-MRI, and P-MRI) demonstrate that ABE-Mamba achieves competitive performance.}, archive = {J_ESWA}, author = {Bingjie Guo and Wenhui Huang and Xiaoyan Wang}, doi = {10.1016/j.eswa.2025.129897}, journal = {Expert Systems with Applications}, month = {3}, pages = {129897}, shortjournal = {Expert Syst. Appl.}, title = {ABE-mamba: Few-shot medical image segmentation via adversarial bidirectional enhanced mamba}, volume = {298}, year = {2026}, }
• (2026). Exact reliability of cold chain networks with multi-state travel time and transport capacity. ESWA, 298, 129892. (www)

  @article{ , abstract = {Post-pandemic lifestyle changes have increased reliance on e-commerce, boosting the logistics sector. One of the most highly regarded industries is cold chain logistics, especially for vaccines and refrigerated foods. In cold chain networks, transport routes have varying capacities based on customer orders, and travel times fluctuate due to traffic and weather. Thus, this study focuses on evaluating network reliability, i.e., the probability to meet given demands within the specified time threshold, of cold chain networks considering the two multi-state factors: travel time and transport capacity. To account for practical situations, a multi-state cold chain network (MCCN) is constructed with retailers, third-party logistics companies, and suppliers as nodes, and transportation routes as arcs. The concept of minimal path is used to determine the transport flow that complies with the time threshold and to determine the transport capacity vectors that satisfy the demands. An algorithm is proposed to resolve different characteristics of time thresholds and demand requirements for efficient assessment. Network reliability is successfully calculated, as shown in the case and sensitivity analysis. This allows managers to grasp the performance of MCCN and make informed decisions based on the achieved network reliability.}, archive = {J_ESWA}, author = {Thi-Phuong Nguyen and Chin-Lung Huang and Louis Cheng-Lu Yeng and Yi-Kuei Lin}, doi = {10.1016/j.eswa.2025.129892}, journal = {Expert Systems with Applications}, month = {3}, pages = {129892}, shortjournal = {Expert Syst. Appl.}, title = {Exact reliability of cold chain networks with multi-state travel time and transport capacity}, volume = {298}, year = {2026}, }
• (2026). Dual-tensor collaborative facilitation for multi-view clustering. ESWA, 298, 129883. (www)

  @article{ , abstract = {In multi-view clustering methods, tensor is often used to capture higher-order information among multi-view data and characterize the feature space structure of the original data. In this paper, we propose the D ual- T ensor C ollaborative F acilitation for M ulti- V iew C lustering approach. DTCF-MVC collaboratively facilitates the multi-view representation tensor through the original data projection tensor, and then finally obtains the multi-view optimal representation through rank-based adaptive weighted fusion. Secondly we design new solutions for tensor rank and tensor sparsity respectively. Therefore, we design a new tensor projection norm to handle the projection tensor. In addition, since there may be large differences between multi-view features, we design a new Weighted Arctangent Coupled tensor nuclear norm. This enables weighted arctangent coupling of singular values of different sizes in the tensor-singular value decomposition, which adequately approximates the true rank of the singular value matrix generated by the representation tensor in the frequency domain decomposition. The superiority of the proposed method is demonstrated by extensive experiments on different datasets.}, archive = {J_ESWA}, author = {Jie Zhang and Xiaoqian Zhang and Yongyi Yang and Jinghao Li and Zhenwen Ren and Rong Tang and Dong Wang}, doi = {10.1016/j.eswa.2025.129883}, journal = {Expert Systems with Applications}, month = {3}, pages = {129883}, shortjournal = {Expert Syst. Appl.}, title = {Dual-tensor collaborative facilitation for multi-view clustering}, volume = {298}, year = {2026}, }
• (2026). PFL-TEnet: Personalized federated learning for wind power forecasting via time-series embedding and hypernetwork modeling. ESWA, 298, 129881. (www)

  @article{ , abstract = {As wind power occupies an increasingly important position in the global energy structure, improving its power prediction accuracy has become the key to guarantee the stable operation of the power system and the efficient consumption of new energy. However, wind energy is driven by complex nonlinear meteorological factors, exhibiting significant time series dependency and heterogeneity among wind farms, while also facing practical requirements for data privacy protection. These factors present numerous challenges for personalized modeling across multiple wind farms. To address these challenges, this paper proposes a personalized federated learning model, PFL-TEnet, which utilizes time-series embedding network to extract critical temporal features and dynamically generates personalized model parameters through hypernetwork. Under the premise of protecting data privacy, this model balances personalized adaptability and generalization capability. Experimental results demonstrate that the proposed method significantly outperforms existing mainstream methods on multiple real wind farm datasets and demonstrates notable advantages in prediction accuracy and personalized performance.}, archive = {J_ESWA}, author = {Yi Li and Chenguang Zhang and Yan Zhang and Jianming Hu}, doi = {10.1016/j.eswa.2025.129881}, journal = {Expert Systems with Applications}, month = {3}, pages = {129881}, shortjournal = {Expert Syst. Appl.}, title = {PFL-TEnet: Personalized federated learning for wind power forecasting via time-series embedding and hypernetwork modeling}, volume = {298}, year = {2026}, }
• (2026). Edge-guided semi-supervised 6D pose estimation with cross-domain alignment for robotic grasping. ESWA, 298, 129880. (www)

  @article{ , abstract = {Accurately estimating 6D poses of objects is crucial for robotic grasping, and existing pose estimation methods generally require large amount of labeled data from same domain, which is difficult for real-world cases. Although some feature transfer methods have been proposed, the gap between source and target scenes still exists. To solve these limitations, we construct a virtual environment with scanned object models and further propose a novel edge-guided semi-supervised 6D pose estimation approach for robotic grasping, where labeled virtual images and unlabeled real images are required, substantially reducing reliance on annotated real data. The designed edge guidance module is able to highlight fine-grained edges of objects and weaken their surface textures, effectively mitigating negative transfer. In addition, a simple yet effective domain mixture mechanism is introduced to align pose-specific features of virtual and real domains from implicit and explicit perspectives, facilitating smoother feature alignment transitions. Comparative experiments are extensively conducted on two pairs of datasets, and the results indicate that the proposed method outperforms other approaches and transfer strategies by 4.1 % and 7.2 %, respectively. The real-world pose estimation and robotic grasping experiments also prove its effectiveness in diverse scenarios.}, archive = {J_ESWA}, author = {Lu Chen and Jiajie Wen and Yan Gao and Jing Yang and Chao Zhang and Shan An and Peng Wu}, doi = {10.1016/j.eswa.2025.129880}, journal = {Expert Systems with Applications}, month = {3}, pages = {129880}, shortjournal = {Expert Syst. Appl.}, title = {Edge-guided semi-supervised 6D pose estimation with cross-domain alignment for robotic grasping}, volume = {298}, year = {2026}, }
• (2026). Multi-source multi-label feature selection with missing features. ESWA, 298, 129879. (www)

  @article{ , abstract = {Feature dimensionality reduction on Multi-Source Multi-Label (MSML) data is a critical and challenging task. Because practical situations always produce massive MSML data, but they usually contain more missing feature values in the high-dimensional feature space and present severely skewed label distributions in the multi-label space, which aggravate the difficulties in the tackling of high-dimensional feature selection on MSML data. However, much attention in feature selection has been directed either toward multi-label data or multi-source data, while little attention is focused on MSML data, not to mention those containing missing features. Motivated by this, we present a new feature selection method for MSML data with missing features, called MMFSMF. Specifically, to overcome the issue of feature missing, we first supplement the feature matrix by constructing a feature correlation matrix during the modeling process of multi-label learning. At the meanwhile, we utilize a multi-label oversampling mechanism to address the persistent problem of label skewness in multi-label data. Secondly, in terms of the above processing, we introduce a refined infinite feature selection algorithm to perform feature dimensionality reduction in each multi-label data source, considering both label correlations and label-specific features. Thirdly, to address feature redundancy among multiple data sources, we apply a new inter-source feature fusion method. Finally, experiments conducted on nine synthetic MSML datasets with missing features demonstrate that MMFSMF achieves superior performances compared to all competing ones.}, archive = {J_ESWA}, author = {Yabo Shi and Peipei Li and Xiulan Yuan and You Wu and Haiping Wang}, doi = {10.1016/j.eswa.2025.129879}, journal = {Expert Systems with Applications}, month = {3}, pages = {129879}, shortjournal = {Expert Syst. Appl.}, title = {Multi-source multi-label feature selection with missing features}, volume = {298}, year = {2026}, }
• (2026). An RL-NSGA-DP algorithm for optimization of robot placement and trajectory allocation in mobile robotic grinding of wind turbine blades. ESWA, 298, 129876. (www)

  @article{ , abstract = {Mobile robot machining, offering a more flexible and reconfigurable approach compared to fixed-base robots, has therefore become a promising solution for efficiently machining large and complex wind turbine blades. In this context, determining proper machining placements and allocating machining trajectories are two pivotal factors in the mobile robotic automation grinding of wind turbine blades, directly affecting machining efficiency and quality. However, the highly nonlinear performance distribution of the robot in the task space, combined with the complexity of the machining surface, presents significant challenges. To address these challenges, this paper presents a general optimization model of this problem with the objectives of completion time and robot manipulability, considering singularity avoidance and collision avoidance. Based on this model, an improved non-dominated sorting genetic algorithm integrated with reinforcement learning and dual population co-evolution (RL-NSGA-DP) is developed. In RL-NSGA-DP, each solution is coded using a novel two-layer metavariable encoding scheme, and a tailored dominated-recessive crossover operator is designed. Moreover, a dual-population collaborative search strategy employing different operators and an adaptive switching environmental selection mechanism based on reinforcement learning are implemented to ensure the convergence and maintain population diversity. Comparative experiments on test instances and a practical case study demonstrate that RL-NSGA-DP outperforms five well-known multi-objective evolutionary algorithms, and effectively addresses robot placement and trajectory allocation problem in mobile robotic machining systems.}, archive = {J_ESWA}, author = {Yi Hua and Xuewu Wang}, doi = {10.1016/j.eswa.2025.129876}, journal = {Expert Systems with Applications}, month = {3}, pages = {129876}, shortjournal = {Expert Syst. Appl.}, title = {An RL-NSGA-DP algorithm for optimization of robot placement and trajectory allocation in mobile robotic grinding of wind turbine blades}, volume = {298}, year = {2026}, }
• (2026). Adaptive GNSS-5G hybrid positioning based on time offset optimization estimation and multi-rate measurement fusion. ESWA, 298, 129875. (www)

  @article{ , abstract = {The fusion positioning of the Global Navigation Satellite System and Fifth-Generation Mobile Communication Network is a key direction for breaking through the performance bottleneck of a single system. However, it faces two core challenges: inconsistent spatiotemporal benchmarks and multi-rate measurement fusion. To address these issues, this paper proposes a joint time offset estimation and phased fusion strategy: An adaptive time-varying offset model is established, and an adaptive relative time offset estimation algorithm based on pseudo-measurements is designed. The high-precision time benchmark provided by GNSS is used to realize the indirect estimation of the 5G absolute time offset, solving the problem of offset accumulation in dynamic scenarios. A two-stage filtering framework is proposed, which processes the coordinate conversion error of 5G polar coordinate measurements through a modified unbiased converted measurement Kalman filter, combines with a Kalman filter to estimate the target state, and constructs a time offset pseudo-measurement based on velocity estimation for efficient solutions. A phased multi-rate fusion strategy is designed: At GNSS sampling moments, adaptive weighted fusion is used to correct the accumulated errors of 5G high-frequency data; at non-GNSS moments, 5G high-frequency measurements and motion state equation predictions are used to maintain tracking accuracy for high-dynamic targets. Simulation results show that the proposed algorithm significantly outperforms eight mainstream algorithms such as SPP, EKF, and UKF in positioning accuracy, with a total average error of 1.66 m and a total root mean square error of 2.02 m. Moreover, the error distribution is more concentrated and stability is stronger, which can effectively adapt to the needs of high-dynamic scenarios and provide reliable solutions for GNSS-5G hybrid positioning.}, archive = {J_ESWA}, author = {Huiyu Chen and Yu Lu and Yao Xing and Xu Zhang and Jiongqi Wang}, doi = {10.1016/j.eswa.2025.129875}, journal = {Expert Systems with Applications}, month = {3}, pages = {129875}, shortjournal = {Expert Syst. Appl.}, title = {Adaptive GNSS-5G hybrid positioning based on time offset optimization estimation and multi-rate measurement fusion}, volume = {298}, year = {2026}, }
• (2026). RAFN: A risk-aware feature network for identifying risk factors in supply chain finance. ESWA, 298, 129874. (www)

  @article{ , abstract = {As supply chain finance businesses expand, traditional risk assessment systems, which rely heavily on manual processes and static rule-based frameworks, are increasingly unable to keep up with the complexity and dynamism of modern risk patterns. This often leads to delayed responses and inefficiencies in risk management. To address key challenges such as difficulties in integrating heterogeneous data, low detection rates for hidden risks, and limited ability to capture dynamic risk patterns, this paper introduces a novel Risk-Aware Feature Network (RAFN) driven by an adaptive attention mechanism. The RAFN model is designed with a dual-channel architecture to process numerical and categorical data separately, employs gated linear units to dynamically merge heterogeneous data streams, and incorporates a multi-head attention mechanism with dynamic coefficients to focus on risk-sensitive features adaptively. Experiments conducted on both public and proprietary datasets show that RAFN outperforms mainstream algorithms, achieving a 1.73%-5.81% improvement in accuracy, recall, and F1-score, while maintaining a strong balance between specificity and recall. Furthermore, this study proposes a closed-loop risk management framework based on RAFN, which integrates “smart contract triggering, off-chain model evaluation, and on-chain consensus validation.” This approach offers an efficient technical solution to break down data silos and enhance the precision of risk identification in supply chain finance, paving the way for more effective and reliable risk control systems.}, archive = {J_ESWA}, author = {Yang Zhang and Yating Zhao and Wenjuan Lian and Bin Jia}, doi = {10.1016/j.eswa.2025.129874}, journal = {Expert Systems with Applications}, month = {3}, pages = {129874}, shortjournal = {Expert Syst. Appl.}, title = {RAFN: A risk-aware feature network for identifying risk factors in supply chain finance}, volume = {298}, year = {2026}, }
• (2026). Task-free continual generative modelling via dynamic teacher-student framework. ESWA, 298, 129873. (www)

  @article{ , abstract = {Continually learning and acquiring new concepts from a dynamically changing environment is an important requirement for an artificial intelligence system. However, most existing deep learning methods fail to achieve this goal and suffer from significant performance degeneration under continual learning. We propose a new unsupervised continual learning framework combining Long- and Short-Term Memory management for training deep learning generative models. The former memory system employs a dynamic expansion model (Teacher), while the latter uses a fixed-capacity memory buffer to store the update-to-date information. A novel Teacher model expansion approach, called the Knowledge Incremental Assimilation Mechanism (KIAM) is proposed. KIAM evaluates the probabilistic distance between the already accumulated information and that from the Short Term Memory (STM). The proposed KIAM adaptively expands the Teacher’s capacity and promotes knowledge diversity among the Teacher’s experts. As Teacher experts, we consider generative deep learning models such as : the Variational Autocencoder (VAE), the Generative Adversarial Network (GAN) or the Denoising Diffusion Probabilistic Model (DDPM). We also extend the KIAM-based model to a Teacher-Student framework in which we use a data-free Knowledge Distillation (KD) process to train a VAE-based Student without using any task information. The results on Task Free Continual Learning (TFCL) benchmarks show that the proposed approach outperforms other models.}, archive = {J_ESWA}, author = {Fei Ye and Adrian G. Bors}, doi = {10.1016/j.eswa.2025.129873}, journal = {Expert Systems with Applications}, month = {3}, pages = {129873}, shortjournal = {Expert Syst. Appl.}, title = {Task-free continual generative modelling via dynamic teacher-student framework}, volume = {298}, year = {2026}, }
• (2026). Trend-aware structure relearning framework for water quality prediction with coupled noise governance. ESWA, 298, 129872. (www)

  @article{ , abstract = {Urban river water quality monitoring is an essential initiative for promoting sustainable and green societal development. Water quality prediction, by effectively modeling the spatiotemporal dependencies between different monitoring sites, holds significant practical value for sites where historical data is difficult to obtain. However, despite progress in spatiotemporal (ST) dependency modeling methods, existing models face challenges in accurately capturing the complex ST dependencies of water quality changes due to the coupled effects of numerical and structural noise. To address this, we propose TaSRef, a T rend- a ware S tructure Re learning F ramework for water quality prediction with coupled noise governance. Specifically, we first separate structural noise and relearn the water topology by analyzing the trend features and relationships of neighboring historical water quality data. Then, based on the relearned topologies and the original noisy data, we thoroughly extract spatiotemporal features to isolate numerical noise. Finally, we achieve robust prediction of water quality in complex environments by aligning and aggregating the trend and ST features of corresponding nodes. Validation on three real datasets shows that the proposed TaSRef framework significantly outperforms baseline models in terms of R 2 , MAE, and RMSE. Through case visualization, results further indicate that TaSRef effectively infers the actual connectivity of river systems and demonstrates strong adaptability in handling coupled noises, thereby validating the robustness and effectiveness of TaSRef for accurate water quality prediction in complex environments.}, archive = {J_ESWA}, author = {Shuo Tong and Yuyang Xu and Jianming Sun and Fuzhen Zhuang and Jian Wu and Guangdi Chen and Haochao Ying}, doi = {10.1016/j.eswa.2025.129872}, journal = {Expert Systems with Applications}, month = {3}, pages = {129872}, shortjournal = {Expert Syst. Appl.}, title = {Trend-aware structure relearning framework for water quality prediction with coupled noise governance}, volume = {298}, year = {2026}, }
• (2026). Novel design and speed-adaptive control of a cable-driven parallel elastic hip exoskeleton for compliant locomotion assistance. ESWA, 298, 129871. (www)

  @article{ , abstract = {Robotic hip exoskeletons hold enormous potential to enhance human locomotion. However, the rigid structures and predefined control laws limit their compliance and adaptability during dynamic human-robot interactions. Here, a novel parallel elastic hip exoskeleton is developed for human locomotion assistance. The exoskeleton utilizes a remote cable actuation system to improve compliance and incorporates a parallel elastic mechanism at the hip wearable components to enhance actuator energy efficiency by generating a compensatory torque. For exoskeleton control, a speed-adaptive torque control strategy is implemented to modulate the assistance torque in real time, based on the user’s gait phase and hip movement frequency estimated by adaptive oscillators. The system was tested on seven healthy subjects, and preliminary results indicate that the parallel elastic element achieves a 40.2 % reduction in peak motor torque through energy conversion. The controller exhibits excellent torque tracking performance and effectively extracts human gait features across walking speeds with hip frequency correlation ( R 2 = 0.89). Furthermore, the hip exoskeleton significantly reduced users’ peak hip moments and muscle activity while preserving natural kinematics. The parallel elastic hip exoskeleton demonstrates strong adaptive assistive capabilities and is expected to enhance locomotion in real-world applications.}, archive = {J_ESWA}, author = {Jing Zhang and Aibin Zhu and Bingsheng Bao and Xinyu Wu and Chunli Zheng and Meng Li and Jing Wang and Yu Zhang and Xue Wu and Xiao Li}, doi = {10.1016/j.eswa.2025.129871}, journal = {Expert Systems with Applications}, month = {3}, pages = {129871}, shortjournal = {Expert Syst. Appl.}, title = {Novel design and speed-adaptive control of a cable-driven parallel elastic hip exoskeleton for compliant locomotion assistance}, volume = {298}, year = {2026}, }
• (2026). Short-term high-speed rail passenger flow forecasting integrated extended empirical mode decomposition with multivariate and bidirectional support vector machine. ESWA, 298, 129870. (www)

  @article{ , abstract = {High-speed rail (HSR) short-term passenger flow forecasting is of great significance for dynamically adjusting operation plans and optimizing transportation resource allocation. For this reason, this paper proposes an innovative complete ensemble empirical mode decomposition with adaptive noise integrated with multivariate and bidirectional support vector machine (CEEMDAN-MBSVM) method with four key steps. First, we analyze the correlations between multiple origin–destination (OD) passenger flows and select strongly correlated ODs incorporated with their opposite OD for joint bidirectional forecasting. Second, we decompose the original passenger flow time series by using period division technique of CEEMDAN, which yield multiple intrinsic mode functions (IMFs) and a residual trend term (RES). Then we apply MBSVM to predict the IMFs of each OD and use trend extrapolation to forecast the RES. Finally, we reconstruct the predicted IMFs and RES to obtain the final bidirectional HSR OD daily passenger flows. Subsequently, we conduct a comprehensive validation exercise and significance testing, using real data from Beijing-Shanghai HSR Line, against seven prediction methods. In particular, for five selected ODs, benchmarking against EEMD-MSVM method, the best performer among the six existing models, our model reduces the minimum mean absolute percentage error (MAPE) by 1.30 % to 4.97 % and benchmarking against ARIMA model, the worst performer among the six existing models, our model reduces the MAPE by 11.57 % to 22.72 %. This research has clearly demonstrated the value of leveraging bidirectional OD data on improving short-term passenger flow forecasting.}, archive = {J_ESWA}, author = {Xueyi Guan and Michael Z.F. Li and Jin Qin and Chengna Wang}, doi = {10.1016/j.eswa.2025.129870}, journal = {Expert Systems with Applications}, month = {3}, pages = {129870}, shortjournal = {Expert Syst. Appl.}, title = {Short-term high-speed rail passenger flow forecasting integrated extended empirical mode decomposition with multivariate and bidirectional support vector machine}, volume = {298}, year = {2026}, }
• (2026). Transfer learning based standard-essential patent prediction with prior transfer direction learning. ESWA, 298, 129869. (www)

  @article{ , abstract = {Standard-Essential Patent Prediction (SEPP) holds strategic significance for technological development and international market competition. Traditional SEPP models learned from Standard-Essential Patents (SEPs) with country-specific distribution differences result in different prediction accuracy. Therefore, we propose two propositions: (1) Can transfer learning be leveraged to improve prediction performance of lower-accuracy countries. (2) Can different transfer directions achieve different transfer learning performances. To address these, we propose a transfer learning based SEPP ewith prior transfer direction learning (TLSEPP-PTDL) model. The model uses a mixed transfer learning method, achieving an average accuracy of 92.03 % on four datasets, surpassing the state-of-the-art (SOTA) by 2.03 % and improving precision, recall, and F1-score by 4.25 %, 0.33 %, and 2.25 %, respectively. Moreover, we conduct experiments across countries with different patent volume, standardization rate, and standardization speed, resulting in positive transfer when transfer learning uses source domains with (1) high volume, high rate, and high speed; (2) high volume, low rate, and high speed; (3) low volume, high rate, and high speed.}, archive = {J_ESWA}, author = {Weidong Liu and Xiaoyu Fan and Kai Wang and Hongjun Sun and Keqin Gan and Cuicui Jiang and Fangyuan Lei}, doi = {10.1016/j.eswa.2025.129869}, journal = {Expert Systems with Applications}, month = {3}, pages = {129869}, shortjournal = {Expert Syst. Appl.}, title = {Transfer learning based standard-essential patent prediction with prior transfer direction learning}, volume = {298}, year = {2026}, }
• (2026). MetaPlanner: A decentralized metaheuristic-driven framework for spatio-temporal trajectory planning of agent swarms in dynamic environments. ESWA, 298, 129868. (www)

  @article{ , abstract = {This paper presents MetaPlanner, a decentralized metaheuristic-driven framework for joint spatio-temporal trajectory planning of multi-agent swarms in dynamic environments. MetaPlanner generates piecewise polynomial spline trajectories by directly optimizing parameterized spatio-temporal control points, innovatively unifies optimization across static and dynamic obstacles, inter-agent collisions, and dynamic constraints, while also supporting dynamic target tracking and precise time-of-arrival control. For swarm coordination, MetaPlanner introduces an efficient decentralized conflict resolution strategy based on a conflict graph and approximate minimum vertex cover. Extensive simulations, including coordinated navigation of up to 100 agents in complex 3D dynamic environments, precise time-of-arrival tasks, dynamic target tracking, and large-scale robustness tests (MetaPlanner-ABC achieved zero collisions in 40,000 experiments), validate the framework’s high levels of safety, coordination, efficiency, and scalability. MetaPlanner’s code and experimental environments will be open-sourced on GitHub at https://github.com/Bziyue/MetaPlanner to foster further research and applications in multi-agent trajectory planning.}, archive = {J_ESWA}, author = {Yaxin Li and Deping Zhang and Yu Wang}, doi = {10.1016/j.eswa.2025.129868}, journal = {Expert Systems with Applications}, month = {3}, pages = {129868}, shortjournal = {Expert Syst. Appl.}, title = {MetaPlanner: A decentralized metaheuristic-driven framework for spatio-temporal trajectory planning of agent swarms in dynamic environments}, volume = {298}, year = {2026}, }
• (2026). IMPACT-net: An integrated multi-scale and computation-efficient timely network for surface defect detection in industrial embedded systems. ESWA, 298, 129867. (www)

  @article{ , abstract = {Automated defect detection is crucial in industrial production, with typical scenarios in steel, automotive, and new energy manufacturing. Although deep learning-based defect detection methods have achieved significant progress, challenges such as limited detection accuracy for low-contrast defects and difficulties in efficient inference still persist. To address these challenges, this paper proposes an integrated multi-scale and computation-efficient timely network (IMPACT-Net) for surface defect detection. Firstly, a Precision-Enhanced Feature Pyramid Network (PE-FPN) is designed to improve the detection performance for low-contrast and fine defects by enhancing multi-scale feature fusion. Secondly, an Adaptive Normalized Wasserstein Distance Loss (ANWD-Loss) is proposed to optimize bounding box localization accuracy and enhance robustness. Finally, by employing Progressive Block-Freezing Architecture Search (PBF-AS) and a ZYNQ-based acceleration platform, computational complexity is significantly reduced, and efficient inference is achieved under low-power conditions. Experimental results show that the proposed IMPACT-Net achieves an mAP50 of 78.9 % on the NEU-DET dataset with 2.3 ms inference time and 71.4 % mAP50 on the GC10-DET dataset with 2.5 ms inference time, demonstrating a good balance between detection accuracy and real-time performance that is well suited for resource-constrained embedded industrial environments.}, archive = {J_ESWA}, author = {Ruiqi Wu and Yong Zhang and Rukai Lan and Lei Zhou}, doi = {10.1016/j.eswa.2025.129867}, journal = {Expert Systems with Applications}, month = {3}, pages = {129867}, shortjournal = {Expert Syst. Appl.}, title = {IMPACT-net: An integrated multi-scale and computation-efficient timely network for surface defect detection in industrial embedded systems}, volume = {298}, year = {2026}, }
• (2026). Frequency-decomposed attention joint optimization network for image compressive sensing. ESWA, 298, 129866. (www)

  @article{ , abstract = {Network approaches for image compressive sensing (ICS) have garnered significant attention due to their high efficiency and fidelity in image reconstruction. Reconstructing complex image textures from highly compressed measurements has been a longstanding goal of ICS, yet existing methods often struggle to varying degrees with the restoration of low-frequency (LF) textures and high-frequency (HF) details, which potentially limits the quality of the reconstructed image. In this paper, we propose a Frequency-decomposed Attention Joint Optimization Network (FAJO-Net) for ICS, which is capable of enhancing the attention to LF and HF components of images. Specifically, we introduce a frequency-decomposed sparse prior and coupling fidelity constraints, and incorporate a tri-optimization network framework for full, low, and high-frequency (FLH) features, where each component is optimized using an optimization-unfolded multi-scale network (OM-Net), inclusive of Principal Component Augmented Gradient Descent Module (PCAGDM) and U-shaped Proximal Mapping Module (UPMM). The PCAGDM optimizes the FLH features efficiently by supplementing the optimization of the minimum dimension principal component augmented features while optimizing the principal component features. The UPMM is able to perform multi-scale proximal mapping for all FLH features. Finally, we design a Frequency-decomposed Interaction Attention Module (FIAM) to enhance the fusion of FLH features, particularly the HF and LF components related to the full-frequency features, while reducing the impact of unnecessary features introduced by frequency decomposition. Extensive experiments demonstrate that our proposed FAJO-Net surpasses the state-of-the-art ICS networks in terms of image fidelity and visual effect, and validates that the proposed FAJO-Net framework can help enhance the image reconstruction capabilities of the vast majority of existing ICS networks, further unlocking the potential for high-fidelity restoration in ICS. Code is available at https://github.com/giant-pandada/FAJO-Net .}, archive = {J_ESWA}, author = {Zhifu Tian and Tao Hu and Di Wu and Shu Wang and Tingli Li and Ming Zhang}, doi = {10.1016/j.eswa.2025.129866}, journal = {Expert Systems with Applications}, month = {3}, pages = {129866}, shortjournal = {Expert Syst. Appl.}, title = {Frequency-decomposed attention joint optimization network for image compressive sensing}, volume = {298}, year = {2026}, }
• (2026). CARNet: Cross-attention guided feature reconstruction for RGB-T object detection. ESWA, 298, 129865. (www)

  @article{ , abstract = {RGB-T object detection leverages the complementary advantages of visible and infrared images in describing the same scene, providing a robust solution for all-weather intelligent perception systems. However, existing methods still suffer from insufficient feature fusion due to inherent differences between the two modalities. To address this issue, we propose CARNet, a cross-attention guided feature reconstruction framework for RGB-T object detection. The core innovation lies in the introduction of a novel cross-attention guided feature reconstruction (CAR) module, which establishes a new paradigm of feature decomposition-interaction-reconstruction for multispectral feature interaction: the feature decomposition unit (FDU) separates strong/weak features based on group normalization weights, the inter-modality iterative cross-attention (Inter-ICA) module enables information interaction from strong to weak features, and the feature reconstruction unit (FRU) reconstructs the modality. This interaction strategy effectively retains modality-specific information while avoiding feature aliasing. Additionally, a channel attention fusion (CAF) module is introduced to optimize the integration of dual-modality features through adaptive weight allocation. Extensive experiments on the KAIST, M3FD and VEDAI public datasets show that CARNet significantly outperforms existing state-of-the-art methods, achieving precise and stable detection of objects in complex scenarios.}, archive = {J_ESWA}, author = {Sirui Wang and Guiling Sun and Liang Dong and Bowen Zheng}, doi = {10.1016/j.eswa.2025.129865}, journal = {Expert Systems with Applications}, month = {3}, pages = {129865}, shortjournal = {Expert Syst. Appl.}, title = {CARNet: Cross-attention guided feature reconstruction for RGB-T object detection}, volume = {298}, year = {2026}, }
• (2026). Identification of critical nodes by fusing propagation probabilities and entropy in binary networks. ESWA, 298, 129861. (www)

  @article{ , abstract = {Identification of critical nodes is crucial for effectively allocating resources and prioritizing tasks in complex networks, which significantly enhances the stability and the efficiency of networks in real-world environments. Generally, existing studies primarily focus on extracting multiple different influential factors from network topology, but they have to face accuracy limitations due to high computational complexity, overlapping influence ranges, and information loss. Inspired by information entropy, in this paper, we explore to identify critical node in complex networks from the perspective of inter-node propagation probabilities. We introduce an innovative critical node ranking algorithm, named MNIE (Mixed Node Information Entropy). MNIE initially segments the node influence within the network topology by distinguishing between global and local effects so as to integrate a more comprehensive topological features set. Then, we refine the connection probability calculation and integrate the features derived from the network structural topology with the probabilities of information transmission (infection rates) among the nodes. Experimental results on 9 real-world networks and 4 synthetic datasets indicate that MNIE enhances the identification of critical nodes and accomplishes better than state-of-the-art methods on monotonicity and accuracy.}, archive = {J_ESWA}, author = {Lintao Zhang and Jianing Zhang and Rong Yan and Guoqin Yu}, doi = {10.1016/j.eswa.2025.129861}, journal = {Expert Systems with Applications}, month = {3}, pages = {129861}, shortjournal = {Expert Syst. Appl.}, title = {Identification of critical nodes by fusing propagation probabilities and entropy in binary networks}, volume = {298}, year = {2026}, }
• (2026). Adaptive energy management for battery swapping stations using HMDE-PSO: Optimizing charge-discharge control against cyber-physical attacks. ESWA, 298, 129860. (www)

  @article{ , abstract = {Battery Swapping Stations (BSSs) are emerging as critical components in smart power systems, offering rapid energy refueling, grid load balancing, and improved battery lifecycle management for electric vehicles (EVs). However, the economic operation and cyber-physical security of BSSs remain underexplored, particularly in microgrids that integrate distributed generation (DG) and face increasing vulnerability to cyber-attacks. This paper presents a novel, adaptive energy management framework that optimally schedules the charge and discharge cycles of BSSs under uncertain EV user behavior and potential cyber-physical disruptions. A key innovation lies in modeling two types of cyber-attacks—power disruption and control hijacking—and embedding their technical and economic impacts directly into the optimization process. To solve this multi-objective problem, a Hybrid multi-objective Differential Evolution–Particle Swarm Optimization (HMDE-PSO) algorithm is proposed, which efficiently balances cost minimization, system reliability, and resilience. The framework is validated using the IEEE 69-bus distribution system, demonstrating substantial improvements: over 40% reduction in power losses, enhanced voltage stability, and lower operational costs compared to conventional methods. This work distinguishes itself by integrating cyber-defense considerations with real-time energy scheduling, providing a comprehensive and resilient solution for future BSS-integrated microgrids.}, archive = {J_ESWA}, author = {Mehdi Ahmadi Jirdehi and Hamdi Abdi and Hazhir Dousti}, doi = {10.1016/j.eswa.2025.129860}, journal = {Expert Systems with Applications}, month = {3}, pages = {129860}, shortjournal = {Expert Syst. Appl.}, title = {Adaptive energy management for battery swapping stations using HMDE-PSO: Optimizing charge-discharge control against cyber-physical attacks}, volume = {298}, year = {2026}, }
• (2026). Attention-enhanced 3D residual networks for knee abnormality classification. ESWA, 298, 129858. (www)

  @article{ , abstract = {The advancement of deep learning technologies, particularly through Convolutional Neural Networks (CNNs), has substantially enriched medical image analysis. This study focuses on improving knee MRI diagnostics by comparing 2D and 3D CNN architectures using the MRNet and SKM-TEA datasets. Initially, modified 2D CNNs, such as ResNet50, were applied for plane-specific and integrated multi-plane analyses. Plane-specific models captured detailed anatomical features, while integrated approaches synthesized information across multiple planes, improving diagnostic capability but lacking full volumetric data utilization. To address these limitations, a novel 3D CNN architecture enhanced with residual attention blocks was developed, leveraging volumetric MRI data. These blocks integrate spatial attention and Squeeze-and-Excitation (SE) mechanisms, optimizing feature focus for accurate diagnostics. This approach improved both model precision and interpretability, which are crucial for clinical applications. Experimental evaluation on the MRNet dataset demonstrated that the proposed 3D CNN outperformed 2D models, achieving 83.58 % accuracy for abnormalities. On the SKM-TEA dataset, the model classified Meniscal Tear (71.36 %), Ligament Tear (79.84 %), Cartilage Lesion (84.28 %), and Effusion (76.74 %), demonstrating robustness in complex pathology detection. Gradient-weighted Class Activation Mapping (Grad-CAM) further enhanced interpretability by highlighting critical diagnostic regions. These findings emphasize the effectiveness of attention-guided 3D CNNs in knee abnormality classification. Future work will explore broader applications in medical imaging, refining the model’s generalizability across diverse clinical datasets.}, archive = {J_ESWA}, author = {Mohamad M.A. Ashames and Semih Ergin and Omer N. Gerek and H. Serhan Yavuz}, doi = {10.1016/j.eswa.2025.129858}, journal = {Expert Systems with Applications}, month = {3}, pages = {129858}, shortjournal = {Expert Syst. Appl.}, title = {Attention-enhanced 3D residual networks for knee abnormality classification}, volume = {298}, year = {2026}, }
• (2026). SmartScope: Smart contract vulnerability detection via heterogeneous graph embedding with local semantic enhancement. ESWA, 298, 129857. (www)

  @article{ , abstract = {Smart contracts are integral to blockchain ecosystems, yet their security remains a critical concern due to the prevalence of exploitable vulnerabilities. Existing conventional and deep learning-based vulnerability detection methods often struggle to capture the fine-grained semantics and heterogeneous structural dependencies essential for accurate analysis. We propose and implement SmartScope , a novel technique for smart contract vulnerability detection that leverages heterogeneous graph embedding with local semantic enhancement. Specifically, SmartScope constructs a semantically rich contract graph that depicts control-flow, data-flow, and fallback relations among critical code elements. To guide the graph learning process, we empirically assign various importance coefficients to vulnerability-relevant subgraphs, thereby enhancing the detection model’s focus on semantically critical regions. The heterogeneous graph transformer is then employed to generate context-aware node representations, which are then passed to an MLP-based detector for vulnerability classification. To the best of our knowledge, this is the first method that structurally encodes domain knowledge into the heterogeneous graph learning for achieving effective smart contract analysis. Experimental results demonstrate that SmartScope outperforms 10 representative conventional and deep learning-based baselines on over 5K smart contracts. The evaluation spans multiple vulnerability types, including reentrancy, timestamp dependence, and infinite loops, highlighting the effectiveness and robustness of our work.}, archive = {J_ESWA}, author = {Zhaoyi Meng and Zexin Zhang and Wansen Wang and Jie Cui and Hong Zhong}, doi = {10.1016/j.eswa.2025.129857}, journal = {Expert Systems with Applications}, month = {3}, pages = {129857}, shortjournal = {Expert Syst. Appl.}, title = {SmartScope: Smart contract vulnerability detection via heterogeneous graph embedding with local semantic enhancement}, volume = {298}, year = {2026}, }
• (2026). Short-term air quality prediction using a multi-scale attention fusion model with 3DIGAT-CBAM-BiLSTM based on spatio-temporal correlation. ESWA, 298, 129856. (www)

  @article{ , abstract = {Air Quality Index (AQI) prediction is crucial for environmental management and public health. However, most existing studies focus on single site modeling, neglecting the complex spatial correlations of meteorological factors and air pollutants. Therefore, a multi-scale spatio-temporal prediction model, 3DIGAT-CBAM-BiLSTM, is proposed to fully capture the spatio-temporal evolution characteristics of AQI. To reduce the interference of redundant information, the Maximum Information Coefficient and Dynamic Time Series Trend Correlation Method are employed to select the neighboring sites and influencing factors that are highly correlated with the AQI of the target site. The original air quality data is decomposed and reconstructed into high-frequency, low-frequency, and trend-term subsequences using Multivariate Variational Mode Decomposition and Sample Entropy to enhance prediction accuracy. To forecast the three-dimensional spatial tensors of these reconstructed subsequences based on time steps, monitoring sites, and influencing factors, we propose the 3DIGAT-CBAM-BiLSTM model. The spatial dependencies between sites are effectively captured by the Improved Graph Attention Network, which constructs a graph adjacency matrix based on MIC and geographic distance. Meanwhile, the Convolutional Block Attention Mechanism enhances the focus on important sites and features by combining channel and spatial attention. Furthermore, the Bidirectional Long Short-Term Memory network extracts global temporal patterns. The experimental results on the Beijing dataset show that the proposed model achieves a relative reduction of 8.53 % in RMSE and 5.83 % in MAE compared with the optimal baseline model, demonstrating clear performance improvements and offering a novel approach for modeling complex spatio-temporal data.}, archive = {J_ESWA}, author = {Liangqiong Zhu and Liren Chen and Huayou Chen}, doi = {10.1016/j.eswa.2025.129856}, journal = {Expert Systems with Applications}, month = {3}, pages = {129856}, shortjournal = {Expert Syst. Appl.}, title = {Short-term air quality prediction using a multi-scale attention fusion model with 3DIGAT-CBAM-BiLSTM based on spatio-temporal correlation}, volume = {298}, year = {2026}, }
• (2026). EEG-based emotion identification from nerve conduction mechanisms: A gustatory-emotion coupling model combined with multiblock attention module. ESWA, 298, 129855. (www)

  @article{ , abstract = {Electroencephalogram (EEG)-based emotion identification enables accurate emotional interaction in brain-computer fusion by decoding brain signals, thereby enhancing the intelligence of human-computer collaboration. Data augmentation (DA) techniques offer a promising solution to the challenge of data scarcity in emotion identification. However, traditional DA methods often overlook the physiological mechanisms underlying EEG data, limiting their effectiveness and constraining the performance of emotion classification. To address this, a DA model based on human nerve conduction mechanisms (NCMs), named the gustatory-emotion coupling model and multiblock attention module (GECM-MBAM), is proposed to improve the performance of emotion identification. First, the 1/ f characteristics and synchronization of brain responses are reproduced in the GECM output when stimulated by EEG. The bionic performance of the model in EEG processing is validated, demonstrating brain-like perception of EEG signals via the GECM. Second, the MBAM is designed based on the characteristics of the GECM output, facilitating data augmentation of emotion-related EEG. Comparative experiments demonstrate that GECM-MBAM remarkably outperforms multiple existing DA models in recognition accuracy ( p < 0.05), confirming its effectiveness and superiority in EEG data augmentation. Finally, when compared with state-of-the-art algorithms and in ablation studies, GECM-MBAM demonstrates superior performance in emotion recognition. Specifically, GECM-MBAM attains accuracies of 96.91 % and 94.52 %, recalls of 96.23 % and 93.86 %, and kappa coefficients of 95.45 % and 94.29 % on the SEED and SEED-IV datasets, respectively. In conclusion, the performance of emotion identification is improved using the GECM-MBAM, offering a novel bionic processing approach for affective computing.}, archive = {J_ESWA}, author = {Wenbo Zheng and Yong Peng and Ancai Zhang and Quan Yuan}, doi = {10.1016/j.eswa.2025.129855}, journal = {Expert Systems with Applications}, month = {3}, pages = {129855}, shortjournal = {Expert Syst. Appl.}, title = {EEG-based emotion identification from nerve conduction mechanisms: A gustatory-emotion coupling model combined with multiblock attention module}, volume = {298}, year = {2026}, }
• (2026). Unifying structural proximity and equivalence for enhanced dynamic network embedding. ESWA, 298, 129854. (www)

  @article{ , abstract = {Dynamic network embedding methods transform nodes in a dynamic network into low-dimensional vectors while preserving network characteristics, facilitating tasks such as node classification and community detection. Several embedding methods have been proposed to capture structural proximity among nodes in a network, where densely connected communities are preserved, while others have been proposed to preserve structural equivalence among nodes, capturing their structural roles regardless of their relative distance in the network. However, most existing methods that aim to preserve both network characteristics mainly focus on static networks and those designed for dynamic networks do not explicitly account for inter-snapshot structural properties. This paper proposes a novel unifying dynamic network embedding method that simultaneously preserves both structural proximity and equivalence while considering inter-snapshot structural relationships in a dynamic network. Specifically, to define structural equivalence in a dynamic network, we use temporal subgraphs, known as dynamic graphlets, to capture how a node’s neighborhood structure evolves over time. We then introduce a temporal-structural random walk to flexibly sample time-respecting sequences of nodes, considering both their temporal proximity and similarity in evolving structures. The proposed method is evaluated using five real-world networks on node classification where it outperforms benchmark methods, showing its effectiveness and flexibility in capturing various aspects of a network.}, archive = {J_ESWA}, author = {Suchanuch Piriyasatit and Chaohao Yuan and Ercan Engin Kuruoglu}, doi = {10.1016/j.eswa.2025.129854}, journal = {Expert Systems with Applications}, month = {3}, pages = {129854}, shortjournal = {Expert Syst. Appl.}, title = {Unifying structural proximity and equivalence for enhanced dynamic network embedding}, volume = {298}, year = {2026}, }
• (2026). CoSemiGNN: Blockchain fraud detection with dynamic graph neural networks based on co-association of semi-supervised. ESWA, 298, 129853. (www)

  @article{ , abstract = {With the development of blockchain technology, the increasing number of cyber frauds has caused huge economic losses, prompting more and more researchers to focus on how to effectively detect criminal activities in the blockchain transaction environment. Currently, graph neural network (GNN)-based methods have made significant progress in the field of blockchain illegal transaction detection due to their advantages in extracting graph structure features. However, existing illegal transaction pattern detection methods usually rely on historical labeled data. In the blockchain transaction environment, transaction data changes over time, and it is often difficult to obtain transaction labels. As a result, the performance of these methods is often unsatisfactory when faced with newly distributed transaction data. To address this challenge, this paper proposes a dynamic graph neural network based on co-association of semi-supervised (CoSemiGNN) for more efficiently identifying illegal transactions in blockchain environments under conditions of dynamically changing transaction data. The model combines semi-supervised learning with a dynamic graph neural network, enabling it to effectively identify novel illegal transaction patterns from unlabeled data and adapt to the evolving blockchain network environment. Specifically, CoSemiGNN captures features of novel transactions by integrating semi supervised learning results. It utilizes co-occurrence relations of edges and co-occurrence feature aggregation of nodes to skillfully integrate semi-supervised methods into feature extraction of transaction graphs, enabling the model to extract novel illegal transaction patterns from unlabeled data. In addition, the model utilizes self attention recurrent neural networks (RNNs) to capture temporal information in transactions, ensuring the dynamics of CoSemiGNN. Finally, we theoretically analyze the model, and experiments on a real Bitcoin transaction dataset demonstrate that CoSemiGNN outperforms existing methods by as much as 30 % in terms of F1 scores for detecting illegal transactions when the transaction data undergoes distributional migration. This research compensates the problem that existing methods ignore the distributional changes of blockchain transaction data, and provides a new perspective and an effective solution for blockchain illegal transaction detection.}, archive = {J_ESWA}, author = {Yulong Wang and Qingxiao Zheng and Xuedong Li and Lingfeng Wang and Ling Lin}, doi = {10.1016/j.eswa.2025.129853}, journal = {Expert Systems with Applications}, month = {3}, pages = {129853}, shortjournal = {Expert Syst. Appl.}, title = {CoSemiGNN: Blockchain fraud detection with dynamic graph neural networks based on co-association of semi-supervised}, volume = {298}, year = {2026}, }
• (2026). Fuzzy prototype transfer learning for non-overlapping cross-domain recommendation. ESWA, 298, 129852. (www)

  @article{ , abstract = {Cross-domain recommendation (CDR) offers an efficient and effective solution to mitigate data sparsity in recommender systems. Existing research primarily focuses on exploring knowledge transfer based on overlapping entities or auxiliary contents between domains. However, there is little research on the real non-overlapping cross-domain recommendation (NCDR) problems, even though it poses a more general and applicable prospect. The core challenge of NCDR lies in the difficulty of finding the correct and useful knowledge transfer bridge between domains without relying on the explicit overlapping identities. Utilizing the inherent similarity and fuzzy characteristics of users and items in the latent feature space, this paper investigates a Fuzzy Prototype Transfer (FPT) learning method for the NCDR problem. FPT jointly optimizes prototypes and individual features for both users and items in target domain under the guidance of source features. An end-to-end learnable fuzzy clustering module based on maximum entropy regularization is proposed to learn both user and item fuzzy clustering assignments and fuzzy fusion prototypes. Lastly, by constructing an asymmetric dual-prototype fuzzy transfer module, similar user and item features across domains are found and aligned effectively. Extensive experiments demonstrate FPT’s superior performance over the state-of-the-art methods while maintaining lower inference and memory costs than those of the baselines.}, archive = {J_ESWA}, author = {Ruxia Liang and Qinglin Huang and Xiaoxuan Shen}, doi = {10.1016/j.eswa.2025.129852}, journal = {Expert Systems with Applications}, month = {3}, pages = {129852}, shortjournal = {Expert Syst. Appl.}, title = {Fuzzy prototype transfer learning for non-overlapping cross-domain recommendation}, volume = {298}, year = {2026}, }
• (2026). LLM-based exploration and analysis of real-time and historical blockchain data. ESWA, 298, 129851. (www)

  @article{ , abstract = {Blockchain technology has revolutionized digital transactions and decentralized applications through its transparent and immutable ledger system, with platforms like Ethereum processing millions of transactions daily. However, as blockchain networks grow, traditional blockchain explorers show limitations when providing intuitive access to this vast data landscape, particularly when handling complex analytical queries, interpreting transaction patterns, and serving users without technical expertise. In this paper, we address these limitations by proposing an intelligent blockchain explorer that combines a Large Language Model (LLM)-powered agent for real-time blockchain interactions with a schema-aware SQL agent for historical data analysis. For real-time interactions, a dedicated blockchain agent connects to live networks through external APIs and specialized tools to process queries about current transactions and network states. When analyzing historical data patterns, we use an approach in which a Retrieval-Augmented Generation (RAG) system enhances the SQL agent’s understanding of the blockchain database schema and structure. This SQL agent subsequently translates natural language queries into SQL commands for efficient data retrieval from our periodically synchronized blockchain database. A query processor, powered by an LLM, intelligently routes user queries between these components based on temporal and contextual requirements, which enables both immediate blockchain state analysis and complex historical data querying. We evaluate our system on diverse blockchain queries, including complex analytical scenarios and multi-step operations. The experimental results demonstrate the effectiveness of our schema-aware SQL agent in accurate query translation and the overall system’s capability in handling both real-time and historical blockchain data exploration tasks.}, archive = {J_ESWA}, author = {S. Gebreab and A. Musamih and K. Salah and R. Jayaraman}, doi = {10.1016/j.eswa.2025.129851}, journal = {Expert Systems with Applications}, month = {3}, pages = {129851}, shortjournal = {Expert Syst. Appl.}, title = {LLM-based exploration and analysis of real-time and historical blockchain data}, volume = {298}, year = {2026}, }
• (2026). RL-DFN: A reinforcement learning-driven dual-view feature fusion network for aspect-based sentiment analysis in online public opinion. ESWA, 298, 129850. (www)

  @article{ , abstract = {Sentiment analysis of online public opinion plays a vital role in understanding emotional dynamics on social media. However, most existing approaches focus on overall sentiment within text, often neglecting fine-grained sentiment information. This limitation restricts the depth of insights obtainable for public opinion monitoring. Aspect-based sentiment analysis (ABSA) addresses this issue by identifying sentiment toward specific aspects. Despite its advantages, current ABSA methods struggle to effectively integrate syntactic and sequential features, leading to incomplete contextual understanding. Additionally, noisy dependency trees introduce irrelevant syntactic features, which mislead sentiment classification and degrade accuracy. To address these challenges, we propose RL-DFN, a reinforcement learning-driven dual-view feature fusion network that enhances the robustness of ABSA models. RL-DFN consists of a dual-view feature fusion network and an actor-critic reinforcement learning module. The dual-view feature fusion network employs a graph attention network (GAT) to extract syntactic features and a transformer to capture sequential features. These complementary features are then fused through a bilinear affine transformation (Biaffine) module which captures fine-grained and bidirectional correlations between syntactic and sequential representations, enabling more expressive cross-view interactions. Simultaneously, an actor-critic reinforcement learning module dynamically refines the dependency tree representations by identifying key dependency types and filtering out noise, ensuring the reliability of syntactic features used in fusion. Extensive experiments on five widely used benchmark datasets demonstrate that RL-DFN significantly outperforms existing models, validating the effectiveness of our approach.}, archive = {J_ESWA}, author = {Ziheng Li and Kui Zhao}, doi = {10.1016/j.eswa.2025.129850}, journal = {Expert Systems with Applications}, month = {3}, pages = {129850}, shortjournal = {Expert Syst. Appl.}, title = {RL-DFN: A reinforcement learning-driven dual-view feature fusion network for aspect-based sentiment analysis in online public opinion}, volume = {298}, year = {2026}, }
• (2026). Truck scheduling optimization at a cold chain cross-docking terminal considering uncertainties and the door-mixed service mode. ESWA, 298, 129849. (www)

  @article{ , abstract = {The increasing global demand for perishable agricultural products necessitates advancements in cold chain logistics. Cross-docking, known for its efficiency, is particularly well-suited for the transfer and distribution of such goods. However, truck scheduling at cold chain cross-dock terminals (CDTs) presents unique challenges, including product perishability, stringent time windows, and temperature-controlled environments. This work investigates a truck scheduling problem within a cold chain CDT, explicitly addressing uncertainties in refrigerated product damage (affecting supply) and repackaging times. A two-stage stochastic programming model is developed to capture these uncertainties. To solve this model, a scenario reduction approach employing K-means++ and K-medoids clustering is used, followed by Sample Average Approximation. Small-scale instances are solved optimally using CPLEX. For larger instances, a novel hybrid heuristic algorithm, combining the global search capabilities of Genetic Algorithms with the local search capabilities of Adaptive Large Neighborhood Search and Simulated Annealing, is proposed. Numerical experiments demonstrate the effectiveness of this algorithm, and sensitivity analysis provides valuable managerial insights.}, archive = {J_ESWA}, author = {Feifeng Zheng and Yuzhi Yi and Ming Liu and Huaxin Qiu}, doi = {10.1016/j.eswa.2025.129849}, journal = {Expert Systems with Applications}, month = {3}, pages = {129849}, shortjournal = {Expert Syst. Appl.}, title = {Truck scheduling optimization at a cold chain cross-docking terminal considering uncertainties and the door-mixed service mode}, volume = {298}, year = {2026}, }
• (2026). A two-stage hybrid heuristic approach combining genetic algorithm and variable neighborhood descent for the clustered electric vehicle routing problem. ESWA, 298, 129848. (www)

  @article{ , abstract = {This paper considers a new variant of the Electric Vehicle Routing Problem (EVRP), termed the Clustered Electric Vehicle Routing Problem (CluEVRP). In CluEVRP, all customers are pre-divided into clusters, and each charging station is either located within a cluster or independent of any cluster. Each electric vehicle must complete service for all customers within the current cluster before proceeding to the next cluster or returning to the depot. Electric vehicles can charge at any available charging station while serving a cluster, but incur a penalty cost upon entering each cluster. The objective is to minimize the total logistics cost, comprising vehicle startup costs, cluster entry penalty costs, and energy consumption costs. To solve CluEVRP, a two-stage hybrid heuristic combining a Genetic Algorithm (GA) and Variable Neighborhood Descent (VND) is proposed (HGA-VND), where GA ensures population diversity and VND enhances local search capability. To evaluate the algorithm’s performance, 75 test instances are adapted from classic Clustered Vehicle Routing Problem (CluVRP) dataset, incorporating electric vehicle characteristics. Computational results demonstrate that HGA-VND consistently obtains high-quality solutions within reasonable time for both CluVRP and CluEVRP instances, exhibiting good performance. Furthermore, sensitivity analysis indicates that moderately increasing vehicle capacity, optimizing battery configuration, and adopting lightweight designs can significantly reduce total operating costs. This study extends traditional EVRP research by introducing clustered customer distribution, enriching solutions for routing problems in practical logistics networks, particularly for “milk run” models in industrial parks, and providing significant managerial insights.}, archive = {J_ESWA}, author = {Yuheng Jin and Xiaoguang Bao and Zhaocai Wang}, doi = {10.1016/j.eswa.2025.129848}, journal = {Expert Systems with Applications}, month = {3}, pages = {129848}, shortjournal = {Expert Syst. Appl.}, title = {A two-stage hybrid heuristic approach combining genetic algorithm and variable neighborhood descent for the clustered electric vehicle routing problem}, volume = {298}, year = {2026}, }
• (2026). Prior feature efficient diffusion model for 3D MRI image super-resolution improving brain connectomics analysis. ESWA, 298, 129847. (www)

  @article{ , abstract = {Magnetic resonance imaging (MRI) is a non-invasive imaging technique used to study the internal structure and function of the human body, especially vital in the frontier exploration of brain science. However, due to hardware limitations and relatively thicker slices, the spatial resolution of MRI image is insufficient to meet the demands of clinical applications. Deep learning-based Super Resolution (SR) methods improve MRI image quality but often result in poor visual quality and over-smoothing. Diffusion-based SR methods can produce complex and perceptually appealing details, but require numerous diffusion steps. Besides, current SR model evaluations rely predominantly on quantitative metrics and visual inspection, yet remain disconnected from clinical integration. To alleviate above problems, this paper presents the Prior Feature Efficient Diffusion Model (PFEDiff) for 3D MRI image SR. To improve the generation efficiency, PFEDiff utilizes Resolution Disparity Maps (RDMs) as the target for the SR task. By incorporating rich prior features from low-resolution MRI through the Prior Information Embedding Module (PIEM), PFEDiff achieves precise SR and strong generalization capabilities. In addition, the designed residual attention noise predictor enables advanced noise prediction, thereby improving the overall performance of the model. The MRI images SR experimental results demonstrate that PFEDiff outperforms SOTA methods, and compared to baseline diffusion models, PFEDiff achieves more than 10 times faster sampling speed. To connect the quantitative metrics and clinical application, we apply PFEDiff to Brain Connectomics Analysis. The results of the brain connectomics experiment confirm the clinical application potential of the proposed method. Our code is publicly available at https://github.com/Rein592977/PFEDiff .}, archive = {J_ESWA}, author = {Renpeng Yao and Jinglong Du and Wenzong Peng and Hongbi Li and Jingbo Zhang and Du Lei and Yuanyuan Jia}, doi = {10.1016/j.eswa.2025.129847}, journal = {Expert Systems with Applications}, month = {3}, pages = {129847}, shortjournal = {Expert Syst. Appl.}, title = {Prior feature efficient diffusion model for 3D MRI image super-resolution improving brain connectomics analysis}, volume = {298}, year = {2026}, }
• (2026). Group-guided prompt learning for vision-language models. ESWA, 298, 129846. (www)

  @article{ , abstract = {Prompt learning has become one of the mainstream approaches for enabling Vision-Language Models (VLMs) to effectively adapt to downstream tasks. Recent approaches enhanced the generalization of models by integrating prior knowledge from large language models (LLMs). However, these approaches overlook the potential value of group knowledge derived from semantic correlations across different classes, which may limit the performance of the model in the face of complex downstream tasks. To overcome this challenge, we propose Group-guided Prompt Learning (GGPL) , which integrates group knowledge into the original text prompts through LLMs. Specifically, GGPL uses LLMs to group all classes and integrates the group knowledge into the original text prompts to construct the final text prompts. Furthermore, we introduce a novel Group Knowledge Alignment (GKA) module, which aligns the learnable prompt features with the pre-trained features that contain group knowledge, preventing the learnable prompt features from feature shift during the training process and thus reducing overfitting. Experimental results across 11 public datasets demonstrate that the proposed GGPL method achieves significant improvement on various prompt learning approaches, while numerous ablation experiments also demonstrate the effectiveness of the each component of our GGPL method.}, archive = {J_ESWA}, author = {Yufei Zheng and Shengsheng Wang and Yansheng Gao}, doi = {10.1016/j.eswa.2025.129846}, journal = {Expert Systems with Applications}, month = {3}, pages = {129846}, shortjournal = {Expert Syst. Appl.}, title = {Group-guided prompt learning for vision-language models}, volume = {298}, year = {2026}, }
• (2026). MTMP: Multimodal targeted molecule generation model with protein features. ESWA, 298, 129845. (www)

  @article{ , abstract = {Molecular generation is a fundamental task in computational chemistry and drug discovery, aiming to design molecular structures with specific properties through algorithmic approaches. Traditional molecular generation models predominantly rely on SMILES representations or molecular topological graphs, limiting their ability to comprehensively capture molecular characteristics. To address this limitation, we propose MTMP, a generative model that bridges molecular structure and sequence. It uses a graph neural network to encode topological features and a GRU decoder to generate SMILES strings. This integration enhances the joint modeling of chemical properties and structure. Furthermore, MTMP incorporates a pre-trained language model trained on large-scale protein sequence data to extract target protein features, enabling a direct and more precise encoding of protein-specific information. This facilitates the generation of molecules with enhanced binding affinity to specific protein targets. To further improve molecular design efficacy, the model was pre-trained on the ZINC molecular database and subsequently fine-tuned via transfer learning using a curated dataset of ligand molecules with known activity against specific target proteins. Experimental evaluations demonstrate that MTMP achieves competitive performance compared to state-of-the-art molecular generation methods. The model produces structurally valid and diverse molecules with favorable physicochemical properties and strong drug-likeness. Notably, it generates novel compounds with high docking scores against target proteins, underscoring its potential for de novo drug design and targeted molecular discovery.}, archive = {J_ESWA}, author = {Dingming Liang and Runfu Yu and Xiaofeng Wang and Kaiyu Dong and Yunjing Zhang and Huicong Liang and Ximing Xu and Tao Song and Shuang Wang}, doi = {10.1016/j.eswa.2025.129845}, journal = {Expert Systems with Applications}, month = {3}, pages = {129845}, shortjournal = {Expert Syst. Appl.}, title = {MTMP: Multimodal targeted molecule generation model with protein features}, volume = {298}, year = {2026}, }
• (2026). Enhancing mixture-of-experts model with prior knowledge for infrared and visible image fusion in complex degraded environments. ESWA, 298, 129844. (www)

  @article{ , abstract = {Infrared and visible image fusion aims to generate a composite image that simultaneously preserves thermal radiation information from infrared images and the rich texture details of visible images. However, existing studies have overlooked the adverse effects of scene degradation in visible images on the fusion process, leading to suboptimal fusion outcomes. To address the challenges posed by scene degradation in image fusion tasks, this paper proposes an image fusion network with degradation correction capability named the Enhancing Mixture-of-Experts model with Prior knowledge for infrared and visible image fusion (EMPFusion), which pioneers the automated execution of multiple degradation restoration tasks during the fusion process. First, we develop a diffusion model for degradation removal to generate high-quality pseudo-labels of visible images, thereby providing supervisory signals for training the fusion network. Second, to overcome the significant challenges in feature extraction caused by complex and diverse degradation scenarios, we design a Degradation removal backbone based on Prior knowledge and the Mixture-of-Experts (DPM) module. This architecture removes degradation with low loss and moderate computational overhead by integrating domain-specific prior knowledge and the Mixture-of-Experts framework. Furthermore, to mitigate semantic loss under extreme environmental conditions, we propose a Semantic Deconstruction and Segmentation (SDS) module based on image-text foundation models, enhancing semantic consistency throughout the fusion process. Extensive experiments demonstrate that EMPFusion excels in infrared-visible fusion tasks within complex degraded scenes. Across the LLVIP, M3FD, RoadScene, and MSRS datasets, EMPFusion achieves state-of-the-art (SOTA) performance on multiple evaluation metrics, showcasing exceptional degradation robustness and visual-semantic information preservation capabilities. By unifying adaptive degradation correction with fusion, this research addresses fusion distortion caused by degraded multimodal data in harsh environments, significantly enhancing applicability and robustness in downstream tasks such as autonomous driving and security monitoring.}, archive = {J_ESWA}, author = {Gang Li and Chengrun Jiang and Jiachen Li and Jin Wan and Mingle Zhou and Delong Han}, doi = {10.1016/j.eswa.2025.129844}, journal = {Expert Systems with Applications}, month = {3}, pages = {129844}, shortjournal = {Expert Syst. Appl.}, title = {Enhancing mixture-of-experts model with prior knowledge for infrared and visible image fusion in complex degraded environments}, volume = {298}, year = {2026}, }
• (2026). Multi-scenario building change detection in remote sensing images using CNN-mamba hybrid network and consistency enhancement learning. ESWA, 298, 129843. (www)

  @article{ , abstract = {Multi-scenario building change detection (BCD) plays a crucial role in scientific decision-making for urban development, natural disasters, and war scenarios. Nevertheless, current research confronts several challenges: (1) In deep learning (DL) methods, convolutional neural networks (CNN) fail to capture global information, and transformer has high computational overhead. (2) The clear distinction and collaborative suppression of scene-level and data-level pseudo-changes remain unresolved. (3) Multi-scenario BCD remains under-explored. To address these challenges, this study proposes a dual pseudo-change suppression framework for multi-scenario BCD. The framework includes the novel C 2 Mamba DL model (CNN collaborates with Mamba) and the consistency enhancement learning strategy (CELS). Among them, the C 2 Mamba merges the local feature extraction capability of CNN with the efficient global modeling capability of Mamba. It further integrates the proposed progressive context information aggregation module (PCIA) and the multi-scale differential feature enhancement module (MDFE). These two modules enable the model to effectively distinguish building changes among various ground targets, thereby suppressing the scene-level pseudo-changes. The CELS first performs data augmentation on the post-temporal images, including weather (clouds, rain, and overcast) and sensor differences. Subsequently, a novel difference-focused loss function is designed to ensure the accurate alignment of the change features between the original and enhanced image pairs, thereby suppressing the data-level pseudo-changes. In experiments, the BCD performance of urban development, natural disasters, and war scenarios is evaluated using WHU-CD, xBD, and WraBCD datasets, respectively. Compared to the second-best methods, the proposed method achieves F1-score improvements of 1.36%, 2.69%, and 1.65%, and IoU improvements of 2.49%, 3.21%, and 2.32%, respectively. Additionally, numerous ablation experiments are conducted to validate the validity of the proposed method. And the robustness of the proposed method is verified through zero-shot generalization and few-shot testing.}, archive = {J_ESWA}, author = {Wei Li and Guorui Ma and Haiming Zhang and Peng Chen and Di Wang and Rong Chen}, doi = {10.1016/j.eswa.2025.129843}, journal = {Expert Systems with Applications}, month = {3}, pages = {129843}, shortjournal = {Expert Syst. Appl.}, title = {Multi-scenario building change detection in remote sensing images using CNN-mamba hybrid network and consistency enhancement learning}, volume = {298}, year = {2026}, }
• (2026). Software defect prediction based on graph code semantics. ESWA, 298, 129842. (www)

  @article{ , abstract = {As software systems become larger and more complicated, researchers are focusing more on how to effectively determine whether a program is flawed. Convolutional neural networks (CNNs) have been employed by several researchers in recent years to extract latent semantic information based on Abstract Syntax Trees (AST). However, the small granularity of code information in AST makes it easy to focus on local information and difficult to capture global semantic information. Further, ASTs lack control flow and data flow edges to fully utilize the code’s contextual semantic information. In this paper, we obtain the Control Flow Graph (CFG) and Program Dependence Graph (PDG) of a program through the improved TinyPDG. The Continuous Bag-of-Words (CBOW) model is utilized to train the corpus of CFG and PDG. Additionally, an improved PNIAT layer that combines multi-head attention and BiLSTM is employed to obtain method-level semantic feature vectors. Subsequently, methods based on weighted summation and linear fully connected are respectively proposed to aggregate method-level semantic feature vectors into file-level semantic feature vectors. The joint features are then constructed in combination with the hand-labeled features. Finally, after balancing the data using the SMOTETomek method, 11 machine learning models are used as classifiers for defect prediction. The experimental findings demonstrate that, compared to current software defect prediction techniques, the PDG using a weighted summation-based approach in the QDA model (W-PDG-QDA) presented in this research is the best, and improves AUC, F1, and accuracy scores by 0.723 % -21.72 %, 5.18 % -55.25 % and 1.99 % -25.16 %.}, archive = {J_ESWA}, author = {Hongwei Tao and Tao Wang and Zhenhao Geng and Xiaoxu Niu and Qiaoling Cao}, doi = {10.1016/j.eswa.2025.129842}, journal = {Expert Systems with Applications}, month = {3}, pages = {129842}, shortjournal = {Expert Syst. Appl.}, title = {Software defect prediction based on graph code semantics}, volume = {298}, year = {2026}, }
• (2026). G-GraphTST: Spatiotemporal learning for multivariate long-term photovoltaic power forecasting. ESWA, 298, 129841. (www)

  @article{ , abstract = {In clean energy systems, the importance of photovoltaic power forecast (PVPF) has become increasingly prominent. In China, the proactive management and control of distributed photovoltaic systems in rural areas have gained policy priority, aiming optimize the integration of photovoltaic energy through multi-node long-term PVPF. However, existing studies face significant challenges in effectively modeling spatiotemporal correlations at different perceptual levels and frequently overlook the interactive learning between spatial and temporal dependencies. To address these limitations, we propose a Global-Embedding Graph-Encoder Time Series Transformer (G-GraphTST 1 ) for solar power forecasting, which features an interactive learning framework comprising spatial and temporal modules to downsample solar data and capture spatiotemporal correlations from multi-level perceptual perspectives. Temporally, the framework compresses high-dimensional raw time-series data into more semantically meaningful 1D local feature representations and introduces learnable global embeddings that interact with all patch embeddings via the self-attention mechanism in the Transformer encoder to capture the periodicity and trends of photovoltaic systems. Spatially, it employs a learnable graph constructor that uses the spatiotemporal heterogeneity of photovoltaic power data in the current batch as query conditions to dynamically reveal the transient power correlations and dependencies between power stations in the photovoltaic network in real time. Extensive experiments on solar datasets and multiple other benchmark datasets demonstrate that GraphTST outperforms existing mainstream methods in performance and exhibits strong generalizability, providing critical support for optimizing smart grid scheduling and managing large-scale photovoltaic power plants.}, archive = {J_ESWA}, author = {Wentao Zheng and Weixing Qian and Pengyu Liu and Rongzhe Wu and Mingjia Chen and Houchen Zhou}, doi = {10.1016/j.eswa.2025.129841}, journal = {Expert Systems with Applications}, month = {3}, pages = {129841}, shortjournal = {Expert Syst. Appl.}, title = {G-GraphTST: Spatiotemporal learning for multivariate long-term photovoltaic power forecasting}, volume = {298}, year = {2026}, }
• (2026). Quantifying expressive power in knowledge graph embeddings: An entropy-based metric framework. ESWA, 298, 129840. (www)

  @article{ , abstract = {Knowledge graph embedding (KGE) is a mainstream technique for the knowledge completion (KGC) task and has attracted significant attention. However, there is a lack of a quantitative evaluation framework to measure the expressive power of KGE models, which is important for model selection and design in downstream applications. Specifically, existing metrics inadequately assess the expressive power of these models because they focus solely on learning outcomes rather than on the embeddings or datasets themselves, which have a significant influence on performance. To address this gap, we propose a novel framework, the Dense Feature Model (DFM), grounded in information theory, to enhance our understanding of KGE. This framework offers three key insights regarding embedding-based KGC (eKGC) approaches: (1) the expressive power of an eKGC model can be quantified using entropy, enabling straightforward model comparison based on a value computed solely from the numbers of entities and relations; (2) the DFM metrics indicate how far an eKGC model is from its theoretical limitations and how much potential improvement remains, as demonstrated by propositions proven in this work; and (3) as a significant byproduct, the framework elucidates why and how a model balances expressive power and computational cost. Furthermore, the framework may be transferable to other relational data processing domains.}, archive = {J_ESWA}, author = {Panfeng Chen and Hui Li and Qi Wang and Xin Zhou and Xibin Wang}, doi = {10.1016/j.eswa.2025.129840}, journal = {Expert Systems with Applications}, month = {3}, pages = {129840}, shortjournal = {Expert Syst. Appl.}, title = {Quantifying expressive power in knowledge graph embeddings: An entropy-based metric framework}, volume = {298}, year = {2026}, }
• (2026). Legend-KINN: A legendre polynomial-based Kolmogorov–Arnold-informed neural network for efficient PDE solving. ESWA, 298, 129839. (www)

  @article{ , abstract = {Physics-Informed Neural Networks (PINNs) have become a widely used method in scientific machine learning for solving partial differential equation (PDE) problems. However, existing PINN methods, while adhering to physical constraints during the convergence process, often suffer from a lack of interpretability in the nonlinear mapping between inputs and outputs during forward propagation. Additionally, fixed activation functions and soft constraint mechanisms result in poor training stability and high computational costs. To address these challenges, we propose a Legendre Polynomial-Based Kolmogorov–Arnold-Informed Neural Network (Legend-KINN), which explicitly enhances physical consistency through Legendre orthogonal polynomials and optimizes the backpropagation path using pseudo-time stepping methods to improve training stability. The use of Legendre polynomials improves the physical interpretability of the model while maintaining strict adherence to physical constraints. Experimental results demonstrate that Legend-KINN outperforms standard Multi-Layer Perceptron (MLPs) and Kolmogorov–Arnold Network (KAN) in terms of both accuracy and computational efficiency, achieving faster convergence with fewer parameters across 25 datasets from three classical computational fluid dynamics cases. Code is available at: https://github.com/zhang-zhuo001/Legend-KINN}, archive = {J_ESWA}, author = {Zhuo Zhang and Xiong Xiong and Sen Zhang and Wei Wang and Yanxu Zhong and Canqun Yang and Xi Yang}, doi = {10.1016/j.eswa.2025.129839}, journal = {Expert Systems with Applications}, month = {3}, pages = {129839}, shortjournal = {Expert Syst. Appl.}, title = {Legend-KINN: A legendre polynomial-based Kolmogorov–Arnold-informed neural network for efficient PDE solving}, volume = {298}, year = {2026}, }
• (2026). A novel grey possibility clustering method based on inverse perspective and its applications. ESWA, 298, 129837. (www)

  @article{ , abstract = {The center-point mixed possibility functions (CMPFs), which are determined by the decision makers qualitatively, are the most important element to obtain the effective clustering results in a grey clustering model. However, different decision makers provide different CMPFs which may lead to inconsistent or contradictory clustering results. In response to this problem, this article proposes an inverse grey possibility clustering model which can determine the CMPFs based on the part of the given final results. This novel matrix-based method can derive all of the required CMPFs which satisfy the partially known clustering results. More specifically, four theorems are put forward to analyze the four different cases, which are single index single object (SISO), single index multiple objects (SIMO), multiple indices single object (MISO) and multiple indices multiple objects (MIMO), respectively, to derive the required CMPFs of a given clustering result using algebraic expressions. For the purpose of developing the matrix representations for the MISO and MIMO situations, a new unified expression of the CMPFs to replace their existing segmented function expression is proposed. Finally, in order to demonstrate how it can be used in practice, the proposed method is applied for evaluating the effects of the reduction of pollution and carbon emissions and determining aerospace equipment component suppliers with different types of data. Compared to the forward GPC models, the proposed IGPC model has higher accuracy.}, archive = {J_ESWA}, author = {Junjie Wang and Xun Li and Yaoguo Dang and Zhongju Shang and Li Ye and Sifeng Liu}, doi = {10.1016/j.eswa.2025.129837}, journal = {Expert Systems with Applications}, month = {3}, pages = {129837}, shortjournal = {Expert Syst. Appl.}, title = {A novel grey possibility clustering method based on inverse perspective and its applications}, volume = {298}, year = {2026}, }
• (2026). Balancing semantic and structural decoding for fMRI-to-image reconstruction. ESWA, 298, 129836. (www)

  @article{ , abstract = {Reconstructing visual images from fMRI signals is an enticing task that opens new horizons in understanding the intricate workings of human cognition. Most existing methods benefit from the diffusion model to decode high-level semantic information from fMRI signals, achieving promising semantic reconstruction. However, such a solution ignores low-level structure information, e.g. , object location and color, leading to an uncompleted visual reconstruction. In this work, we present a novel fMRI-to-image approach to reconstruct high-quality images by balancing semantic and structural decoding in the diffusion model. Specifically, we first utilize the CLIP model and an MLP module to extract sufficient semantic information and structural details, respectively. Then we design a S emantic and S tructural A wareness B alanced module ( SSAB ) to predict the weight of structural information for the current denoising step, thus generating high-quality images by gradually integrating semantic and structural information during image reconstruction. Experimental results demonstrate that the proposed SSAB model is effective with only a few extra parameters, it achieves state-of-the-art performance in comprehensively evaluating both semantic and structural metrics. All code is available on https://github.com/Venchy-he/SSAB .}, archive = {J_ESWA}, author = {Wanqi He and Jin Wang and Hui Li and Hanyang Chi and Bingfeng Zhang}, doi = {10.1016/j.eswa.2025.129836}, journal = {Expert Systems with Applications}, month = {3}, pages = {129836}, shortjournal = {Expert Syst. Appl.}, title = {Balancing semantic and structural decoding for fMRI-to-image reconstruction}, volume = {298}, year = {2026}, }
• (2026). Cohen’s class bilinear distributions and convolutional neural networks applied to broken rotor bar diagnosis. ESWA, 298, 129835. (www)

  @article{ , abstract = {Time-frequency (t-f) signal processing techniques are particularly advantageous for induction motor (IM) fault diagnosis under dynamic and variable industrial operating conditions. Broken rotor bar (BRB) faults remain among the most challenging to detect because of their proximity to the fundamental frequency and significantly lower amplitude in comparison. Additionally, traditional approaches often result in false positives or negatives in scenarios involving load variation, power quality issues, and inverter-fed operations. To address these issues, this work proposes a comprehensive and objective methodology to evaluate eight Cohen-Class Bilinear Distributions (CCBD) to diagnose BRB. CCBDs offer high-resolution t-f representations, a crucial advantage for fault identification. However, their use is limited by cross-terms, nonlinear artifacts inherent to bilinear processing. To overcome this limitation, convolutional neural networks (CNNs) are applied to automatically classify t-f images and identify the CCBD methods that effectively minimize the cross-terms while preserving fault signature harmonics. This strategy also avoids subjective and time-consuming visual inspections. In addition, this work proposes a novel CNN architecture with an attention module (CNN-Attention), designed to enhance performance in this context. The evaluation considers challenging conditions, including 1) line-fed and 2) inverter-fed operation, 3) voltage unbalance, and 4) load oscillations, applied to a 2 HP, 60 Hz motor. Generalization capability is validated with data collected from a different laboratory, using an independent 1 HP, 50 Hz motor and five different inverter models. Experimental results show that combining CNN-Attention with CCBDs enables highly accurate and fast classification, achieving approximately 96% accuracy even when trained and tested on distinct laboratory datasets, demonstrating the effectiveness and adaptability of the proposed method.}, archive = {J_ESWA}, author = {Avyner L.O. Vitor and Alessandro Goedtel and Wesley A. Souza and Marcelo F. Castoldi and Daniel Morinigo-Sotelo and Oscar Duque-Perez}, doi = {10.1016/j.eswa.2025.129835}, journal = {Expert Systems with Applications}, month = {3}, pages = {129835}, shortjournal = {Expert Syst. Appl.}, title = {Cohen’s class bilinear distributions and convolutional neural networks applied to broken rotor bar diagnosis}, volume = {298}, year = {2026}, }
• (2026). A data-driven spatio-temporal driving risk field mechanism for path planning. ESWA, 298, 129834. (www)

  @article{ , abstract = {Characterizing the future risk posed by surrounding human-driven vehicles is crucial for enhancing the safety of autonomous vehicles. Existing risk field methods build spatiotemporal risk fields using mathematical models with fixed parameters, making them struggle to capture dynamic human driving behaviors such as frequent acceleration, deceleration or lane changes, and are prone to overlooking rare but critical sudden events, which leads to unstable risk assessments in complex long-term scenarios. To address the aforementioned issues, a data-driven spatio-temporal risk field framework is proposed, which builds on a Bidirectional Deep Ultra-Gated Recurrent Unit (BDUGRU) to capture the high-dimensional spatio-temporal features of nearby vehicles and precisely predict vehicle distribution patterns over extended horizons. The introduced approach manages to yield a more accurate risk field and significantly improves long-term risk assessment in complex traffic environments. Furthermore, to validate the model’s practicality in engineering, we integrated Rapidly-exploring Random Tree with spatiotemporal data-driven risk field (SRF-RRT) and conducted path-planning simulations for autonomous vehicles using real-world traffic data. The results demonstrate that the proposed model excels in both prediction accuracy and reliability, and effectively reduces the measurement error based on collision time (TTC), offering strong applicability and providing a novel theoretical foundation and technological route for path planning in intelligent connected vehicles (ICVs).}, archive = {J_ESWA}, author = {Zhuoer Wang and Baohan Shi and Jianping Zhang and Xiaowen Zhu and Jian Zhou and Bingrong Xu and Bijun Li}, doi = {10.1016/j.eswa.2025.129834}, journal = {Expert Systems with Applications}, month = {3}, pages = {129834}, shortjournal = {Expert Syst. Appl.}, title = {A data-driven spatio-temporal driving risk field mechanism for path planning}, volume = {298}, year = {2026}, }
• (2026). Automated pothole detection and volume assessment using PointPSSN and smartphone LiDAR point clouds. ESWA, 298, 129833. (www)

  @article{ , abstract = {The rapid detection and assessment of potholes are critical for ensuring road traffic safety. However, point cloud-based techniques relying on surveying vehicles or drones are often expensive and may be limited by roadside obstruction or narrow roadways. This study proposes a novel approach for assessing road potholes using point cloud data collected by smartphone LiDAR. The method integrates the Point Pothole-Specialized Segmentation Network (PointPSSN), a lightweight point cloud segmentation model designed to achieve high accuracy with low parameter complexity and rapid inference, together with scale-adjustable voxelization for assessment. The PointPSSN model incorporates a Geometric Feature Encoder module to capture the geometric attributes of potholes by extracting local geometric features. Neighbor Finder module identifies and aggregates neighboring points that provide more significant information. Experiments were conducted using a smartphone LiDAR device within a 7.28 km 2 area of Wuchang District, Wuhan, China, encompassing diverse road conditions. A dataset of 1040 potholes was constructed for model training and evaluation. The results demonstrate that the PointPSSN model achieves a segmentation accuracy of 97.336 %, precision of 91.322 %, recall of 79.888 %, an F1-score of 85.223 %, and an intersection-over-union (IoU) of 74.251 %. Notably, the accuracy, F1-score, and IoU surpass the performance of state-of-the-art models by 0.233 %, 1.336 %, and 2.006 %, respectively. In terms of efficiency, PointPSSN requires only one-seventh of the FLOPs and one-fifteenth of the parameters of state-of-the-art models, while achieving an 18.37 % faster inference speed. Furthermore, the average relative errors in depth and volume assessment using voxelization methods are 9.08 % and 9.04 %, respectively.}, archive = {J_ESWA}, author = {Tingrui Zhang and Xuequan Zhang and Zichuan Yang and Yumin Chen and Li Song and Weichen Zhang}, doi = {10.1016/j.eswa.2025.129833}, journal = {Expert Systems with Applications}, month = {3}, pages = {129833}, shortjournal = {Expert Syst. Appl.}, title = {Automated pothole detection and volume assessment using PointPSSN and smartphone LiDAR point clouds}, volume = {298}, year = {2026}, }
• (2026). PAD: Popularity-aware debiasing for high-value item recommendation. ESWA, 298, 129830. (www)

  @article{ , abstract = {Recommender systems play a crucial role in our daily lives. However, in the context of high-value item recommendation, they face significant challenges. Due to the high price of these items, user purchase histories are often extremely sparse, making it difficult for recommender systems to accurately capture user preferences. Consequently, they tend to over-rely on popularity information. Moreover, the high-value item market exhibits a pronounced imbalanced distribution, where most user interactions focus on popular items. As a result, traditional recommender systems tend to prioritize these items while rarely recommending less popular ones, leading to low recommendation coverage. To address this challenge, we propose a P opularity- A ware D ebiasing (PAD) model, which improves recommendation coverage in high-value item scenarios without compromising accuracy. First, we employ soft prompts to guide a pre-trained language model (PLM) in enriching user representations. By incorporating semantic knowledge from the PLM, our model captures more comprehensive user preferences, ensuring recommendation accuracy while mitigating the model’s dependence on popularity signals. Building upon this, we apply popularity-aware debiasing to reduce overfitting and enhance coverage. PAD prevents the recommendation model from indiscriminately recommending the most popular items to all users, encouraging it to explore a wider range of items in its recommendations. Experiments conducted on industrial and public datasets demonstrate that our method mitigates popularity bias, significantly improving item recommendation coverage while maintaining accuracy.}, archive = {J_ESWA}, author = {Yuchen Zheng and Dongming Zhao and Xiangrui Cai and Yanlong Wen and Xiaojie Yuan}, doi = {10.1016/j.eswa.2025.129830}, journal = {Expert Systems with Applications}, month = {3}, pages = {129830}, shortjournal = {Expert Syst. Appl.}, title = {PAD: Popularity-aware debiasing for high-value item recommendation}, volume = {298}, year = {2026}, }
• (2026). HWA-net: Hierarchical window aggregate network for cross-resolution remote sensing change detection. ESWA, 298, 129829. (www)

  @article{ , abstract = {Cross-resolution remote sensing change detection (CD) is a critical task in various applications, including urban monitoring, environmental changes, and disaster management, where images captured at different times often possess varying spatial resolutions. Current methods typically address this by resampling low-resolution (LR) images to high-resolution (HR) formats, but such image-level strategies lead to significant artifacts and misalignment in the change map. These imperfections not only reduce detection accuracy but also lead to misleading or false change identifications, resulting in incorrect or incomplete conclusions in time-sensitive applications, such as land-use change detection or disaster monitoring. To address these challenges, we propose the Hierarchical Window Aggregate Network(HWA-Net), a novel framework that directly operates on cross-resolution image pairs without preprocessing, aiming to accurately aggregate cross-resolution representations for robust CD. HWA-Net initially employed window-based feature extraction to produce scale-independent representations, subsequently transferring these features to layered decoding. This process effectively enhances detection accuracy across diverse resolutions. Our approach establishes new state-of-the-art results on three synthesized datasets and one real-world cross-resolution change detection dataset.}, archive = {J_ESWA}, author = {Hualin Yang and Boran Ren and Zhijun Yang and Jing Xiong and Xiying Li and Calvin Yu-Chian Chen}, doi = {10.1016/j.eswa.2025.129829}, journal = {Expert Systems with Applications}, month = {3}, pages = {129829}, shortjournal = {Expert Syst. Appl.}, title = {HWA-net: Hierarchical window aggregate network for cross-resolution remote sensing change detection}, volume = {298}, year = {2026}, }
• (2026). GRDGNN: A directed graph neural network framework for multi-relational inference of gene regulatory networks. ESWA, 298, 129827. (www)

  @article{ , abstract = {Inferring gene regulatory networks (GRNs) from gene expression data is a central challenge in systems biology. Graph neural networks (GNNs) offer a promising approach due to their ability to process graph-structured data. However, existing GNN methods for GRN inference often treat the problem as binary classification, limiting their ability to capture comprehensive regulatory relationships. This paper introduces two learning algorithms that utilize an end-to-end gene regulatory directed graph neural network (GRDGNN) schema for efficient inference of causal relationships in large-scale networks. These algorithms incorporate a directed graph neural network (DGNN) and a graph multi-classification task to identify explicit interactions between transcription factors (TFs) and target genes. The proposed approach consists of four key steps: (1) constructing a directed initial network using regression Pearson correlation and mutual information analysis, (2) extracting subgraphs of observed TF-gene pairs and applying a DGNN for information aggregation, (3) projecting the aggregated information into a low-dimensional space using graph pooling to generate graph representations of TF-gene pairs, and (4) classifying subgraphs using a multilayer perceptron (MLP) for link prediction and inference of explicit regulatory relationships. Evaluation of the DREAM5 microarray and scRNA-seq datasets demonstrates that our transductive and inductive learning methods can accurately and effectively infer explicit regulatory relationships compared to benchmark methods. These results demonstrate that the proposed GRDGNN schema exhibits strong generalization across species, data types, and modalities in cross-species, cross-data type, and cross-modality learning.}, archive = {J_ESWA}, author = {Wanhong Zhang and Zhenyu Guo}, doi = {10.1016/j.eswa.2025.129827}, journal = {Expert Systems with Applications}, month = {3}, pages = {129827}, shortjournal = {Expert Syst. Appl.}, title = {GRDGNN: A directed graph neural network framework for multi-relational inference of gene regulatory networks}, volume = {298}, year = {2026}, }
• (2026). Photovoltaic cluster power ultra-short-term cross-seasonal prediction integrating multi-channel information probabilistic diffusion generation and improved offset loss strategy. ESWA, 298, 129826. (www)

  @article{ , abstract = {Accurate photovoltaic (PV) power prediction under complex meteorological conditions remains challenging, particularly given the pronounced seasonal variations that obscure generation patterns. This study presents a novel ultra-short-term prediction framework integrating meteorological volatility analysis with seasonal characteristic modeling. We developed a specialized multi-channel Gram angular summation field (MGASF) transformation matrix to holistically capture meteorological fluctuations, subsequently leveraging denoising diffusion probabilistic model (DDPM) for strategic augmentation of under-represented weather scenarios to enhance similar-day identification. Our hybrid architecture combines multi-channel vision Transformer (VIT) with bidirectional long and short-term memory (BILSTM) networks to synergistically analyze temporal dependencies and spatial patterns in PV similarity recognition. Furthermore, we engineered a seasonal-adaptive prediction system through an improved variable-weight Smooth L1 loss function, establishing an optimized seasonal alignment mechanism that achieves high-precision prediction across varying meteorological conditions with minimal computational overhead. Through rigorous validation using operational data from a utility-scale photovoltaic cluster in Western Inner Mongolia, the proposed method achieved consistent accuracy improvements: 3.02 % reduction in N RMSE , 1.65 % decrease in N MAE , and 2.19 % enhancement in R 2 compared to baseline approaches in PV cluster. These statistically significant enhancements demonstrate our framework’s capability to mitigate seasonal impacts while maintaining prediction reliability in complex meteorological environments.}, archive = {J_ESWA}, author = {Mao Yang and Yue Jiang and Yunfeng Guo and Jianfeng Che and Wei He and Kang Wu}, doi = {10.1016/j.eswa.2025.129826}, journal = {Expert Systems with Applications}, month = {3}, pages = {129826}, shortjournal = {Expert Syst. Appl.}, title = {Photovoltaic cluster power ultra-short-term cross-seasonal prediction integrating multi-channel information probabilistic diffusion generation and improved offset loss strategy}, volume = {298}, year = {2026}, }
• (2026). Segmentation, fusion, and representation: A novel approach to multi-label classification for long texts. ESWA, 298, 129825. (www)

  @article{ , abstract = {Multi-label text classification (MLTC) is a vital task in natural language processing (NLP), often requiring high-quality text representations generated by pre-trained language models (PLMs). However, the inherent input length constraints of PLMs limit their capacity to handle long texts effectively. To address this challenge, we propose an innovative framework for multi-label long text classification. Our approach incorporates a dynamic text segmentation algorithm that optimally partitions long texts, thereby mitigating the input length limitations of PLMs. Additionally, we enhance both text and label representations by integrating external knowledge, modeling label co-occurrence relationships, and employing attention mechanisms. Extensive experiments conducted on diverse MLTC datasets demonstrate the superior performance of our method and uncover intricate relationships between texts and their associated labels. The code is available at https://github.com/Coder-Jeffrey/SKFRL}, archive = {J_ESWA}, author = {Xin Wang and Junfeng Xiao and Wang Zhang and Tao Deng and Qian Wang}, doi = {10.1016/j.eswa.2025.129825}, journal = {Expert Systems with Applications}, month = {3}, pages = {129825}, shortjournal = {Expert Syst. Appl.}, title = {Segmentation, fusion, and representation: A novel approach to multi-label classification for long texts}, volume = {298}, year = {2026}, }
• (2026). Enhanced multi-scale TCN ensemble for component chain prediction using evolutionary algorithms. ESWA, 298, 129824. (www)

  @article{ , abstract = {In software engineering, component-based development has gained significant traction due to its ability to enhance efficiency and reusability, especially as the complexity of software projects increases. A core aspect of this paradigm is component chain prediction, which intelligently infers subsequent components based on historical usage sequences, effectively reducing redundant development efforts. To overcome limitations in existing methods regarding temporal feature capture and hyperparameter optimization, we propose an Evolution algorithm-enhanced Temporal Convolutional Network ensemble model (E-TCN). This approach integrates the convolutional structure of TCN with the sequential processing capabilities of GRU and employs a dynamic threshold mechanism to improve temporal feature capture and generalization performance. Additionally, we construct a hybrid evolutionary search framework, combining the strengths of four optimization algorithms—Artificial Bee Colony (ABC), Ant Colony Optimization (ACO), Black Widow Optimization Algorithm (BWOA), and Grasshopper Optimization Algorithm (GOA)—for efficient parameter space exploration. Experimental results on real-world development datasets demonstrate a 5.9 % improvement in F1 score compared to conventional TCN models, underscoring the efficacy and practical applicability of the proposed approach in component chain prediction.}, archive = {J_ESWA}, author = {Wen Qin and Peiyang Wei and Tong Peng and Hongping Shu and Zhuoning Zhao}, doi = {10.1016/j.eswa.2025.129824}, journal = {Expert Systems with Applications}, month = {3}, pages = {129824}, shortjournal = {Expert Syst. Appl.}, title = {Enhanced multi-scale TCN ensemble for component chain prediction using evolutionary algorithms}, volume = {298}, year = {2026}, }
• (2026). Towards large-scale multi-objective feature selection: A two-stage evolutionary algorithm guided by dual feature weightings. ESWA, 298, 129823. (www)

  @article{ , abstract = {Feature Selection (FS) is a critical task in high-dimensional data processing, aiming to identify the most discriminative subset of features to improve model performance and reduce computational complexity. In recent years, multi-objective evolutionary algorithms have been widely applied to FS problems due to their ability to simultaneously optimize multiple objectives (i.e., classification accuracy and subset size for an FS problem). However, when dealing with large-scale multi-objective FS problems, existing algorithms often suffer from the vast search space and limited search capability, which makes them prone to local optima. To address these challenges, this paper proposes a two-stage evolutionary algorithm guided by dual feature weightings, named TSEA/DFW. In the first stage, an evolutionary search is performed under the guidance of the filter-based feature weighting strategy. The key features are then identified based on the population distribution and optimal solutions, thereby shrinking the search space. In the second stage, a refined search is conducted in the shrunken feature space to boost search efficiency and solution quality. To this end, a novel weighting strategy named Pareto-based hierarchical feature weighting is proposed, which captures the variation in feature performance across different non-dominated levels, reinforces the contribution of high-quality solutions, and preserves useful information from suboptimal solutions. Additionally, a novel offspring reproduction procedure guided by stage-specific feature weights is designed to further enhance search capability. Experimental results on 13 real-world datasets show that the proposed TSEA/DFW performs best on 10 datasets in terms of HV metric and on 11 datasets in terms of IGD, demonstrating the significant superiority of TSEA/DFW over seven state-of-the-art feature selection methods. The performance improvements stem from the two-stage evolutionary framework guided by dual feature weighting, which enables the early identification of important features, thereby effectively reducing the search space and enhancing search efficiency. In addition, further analysis demonstrates that the proposed TSEA/DFW has strong generality across diverse classifiers, and the developed two-stage evolutionary framework in TSEA/DFW is a general powerful framework that can integrate any mainstream FS algorithm into its second stage, exhibiting robust applicability and scalability.}, archive = {J_ESWA}, author = {Gaohui Li and Zefeng Chen and Yuren Zhou and Zhengxin Huang and Xiaoyun Xia}, doi = {10.1016/j.eswa.2025.129823}, journal = {Expert Systems with Applications}, month = {3}, pages = {129823}, shortjournal = {Expert Syst. Appl.}, title = {Towards large-scale multi-objective feature selection: A two-stage evolutionary algorithm guided by dual feature weightings}, volume = {298}, year = {2026}, }
• (2026). Multi-temporal ensemble for few-shot action recognition. ESWA, 298, 129821. (www)

  @article{ , abstract = {Few-Shot Action Recognition (FSAR) aims to recognize novel action classes with only a few labeled samples. Due to the scarcity of labeled data, FSAR models suffer from high variance and low confidence. To address this issue, this paper first introduces ensemble learning into the field of FSAR, leveraging the diversity among multiple temporal action representations to generate base models. Specifically, we propose a Multi-Temporal Ensemble (MTE) method for FSAR. By combining sub-sequences of video frames of various lengths (i.e., tuples), MTE creates multiple sets of action representations and generates base models based on these representations. All base models share a single embedding network to learn frame-level features. The proposed method adaptively captures temporal relations with different lengths and speeds while avoiding the computational cost of training multiple deep neural networks. Furthermore, we introduce a Short-term Temporal Modeling Module (STMM) that uses self-attention to highlight frames with high variation, enhancing short-term temporal representation at the frame level. The proposed method has been validated on four benchmark datasets. Extensive experimental results demonstrate that MTE outperforms 26 state-of-the-art FSAR methods. The source code is available at https://github.com/CharmainCahill/MTE.git .}, archive = {J_ESWA}, author = {Zhen Jiang and Jianlong Sun and Haodong Liu and Haizhen Guan}, doi = {10.1016/j.eswa.2025.129821}, journal = {Expert Systems with Applications}, month = {3}, pages = {129821}, shortjournal = {Expert Syst. Appl.}, title = {Multi-temporal ensemble for few-shot action recognition}, volume = {298}, year = {2026}, }
• (2026). MSIDiff:Multi-stage interaction-aware diffusion model for protein-specific 3D molecule generation. ESWA, 298, 129820. (www)

  @article{ , abstract = {Structure-based drug design (SBDD) focuses on developing 3D ligand molecules that bind with high affinity to specific protein targets, which requires the accurate capture of the complex interactions between proteins and ligands. Although existing diffusion models have demonstrated potential in molecular generation tasks, they typically consider only a single stage of the generation process. This limitation prevents them from integrating the multi-stage protein-ligand interaction information from both forward and reverse processes, which may negatively impact the binding affinity of the generated molecules. To address this problem, MSIDiff ( M ulti- S tage I nteraction-Aware Diff usion Model), a multi-stage interaction-aware diffusion model for protein-specific molecule generation, is proposed. MSIDiff leverages the pre-trained model MSINet to extract authentic protein-ligand interaction information during the initial diffusion stage and incorporates this information into the reverse process to ensure that the generated molecules accurately interact with target proteins. Through a scoring mechanism, MSIDiff filters key nodes to extract crucial protein-ligand interaction data and employs a GRU-based cross-layer interaction update module to recursively integrate information across different denoising stages, facilitating effective cross-layer information transmission. Experimental results on the CrossDocked2020 dataset show that MSIDiff can generate molecules with more realistic 3D structures and higher binding affinity to protein targets, achieving an Avg. Vina Score of up to -6.36, while maintaining appropriate molecular properties.Our code and data are available at: https://github.com/zhangyaoxiang/MSIDiff .}, archive = {J_ESWA}, author = {Yaoxiang Zhang and Junteng Ma and Ze Zhang and Zhaoyang Dong and Shuang Wang}, doi = {10.1016/j.eswa.2025.129820}, journal = {Expert Systems with Applications}, month = {3}, pages = {129820}, shortjournal = {Expert Syst. Appl.}, title = {MSIDiff:Multi-stage interaction-aware diffusion model for protein-specific 3D molecule generation}, volume = {298}, year = {2026}, }
• (2026). OSATG-GPT: Instruction-tuning large language models with open-source atomic tasks in github. ESWA, 298, 129819. (www)

  @article{ , abstract = {Across numerous application scenarios in Natural Language Processing (NLP), Large Language Models (LLMs) have demonstrated exceptional capabilities in text comprehension and generation. These models exhibit significant potential across various interdisciplinary fields. However, their effectiveness is somewhat constrained by the unique characteristics of the open-source ecosystem. Developing an LLM with generalization capabilities across datasets and tasks, specifically tailored for the open-source ecosystem, is an urgent research need. To address this challenge, this paper introduces open-source atomic tasks, which are defined as intermediate tasks essential for solving complex objectives. These tasks are designed through strategies such as simplification, reversal, decomposition, and composition, enabling models to gradually acquire domain knowledge and understand task interdependencies. By integrating public resources with open-source atomic tasks, we construct OSE-Instruct–an instruction dataset for the open-source ecosystem. We first unify open-source atomic tasks within an instruction-tuning paradigm that reflects real-world developer behavior, and develop OSATG-GPT at various parameter scales by fine-tuning the BLOOMZ backbone model on OSE-Instruct. This enables the model to learn fine-grained developer actions and the underlying task dependencies. Extensive experiments validate the effectiveness of OSATG-GPT compared to other advanced LLMs with larger parameter scales, and highlight its advantages over GPT-4 in specific and complex open-source collaboration tasks.}, archive = {J_ESWA}, author = {Fanyu Han and Li Ma and Fenglin Bi and Yantong Wang and Mingdong You and Wei Wang and Jiaheng Peng and Xiaoya Xia}, doi = {10.1016/j.eswa.2025.129819}, journal = {Expert Systems with Applications}, month = {3}, pages = {129819}, shortjournal = {Expert Syst. Appl.}, title = {OSATG-GPT: Instruction-tuning large language models with open-source atomic tasks in github}, volume = {298}, year = {2026}, }
• (2026). Markov-based continuous learning with diversion of data distribution direction for streaming data in limited memory. ESWA, 298, 129818. (www)

  @article{ , abstract = {Traditional online classifiers often require accumulating past data, leading to uncontrollable memory usage and learning times. The ideal solution is a Markov-based continuous learning approach, where a model updates using only its current state and new data. While one-pass learning with hyper-ellipsoids aligns with this principle, a critical weakness persists: classification ambiguity for data points within the overlap region where ellipsoids from different classes intersect. To solve this, this paper proposes the Diversion of Data Distribution Direction (D 4 ), a new method that implements this Markov-based approach while specifically targeting the ambiguity problem. D 4 introduces two novel mechanisms: a new adaptive width adjustment to prevent over-adjusted ellipsoid boundaries and a distribution diversion technique that resolves ambiguity by projecting data into an optimally selected subspace. The proposed D 4 method was evaluated against seven state-of-the-art online classifiers across nine benchmark datasets, having 2011 to 567,498 samples. It achieved the highest accuracy and macro F1-score on six datasets while proving to be the most computationally efficient and generating the most compact models.}, archive = {J_ESWA}, author = {Peemapat Wongsriphisant and Kitiporn Plaimas and Chidchanok Lursinsap}, doi = {10.1016/j.eswa.2025.129818}, journal = {Expert Systems with Applications}, month = {3}, pages = {129818}, shortjournal = {Expert Syst. Appl.}, title = {Markov-based continuous learning with diversion of data distribution direction for streaming data in limited memory}, volume = {298}, year = {2026}, }
• (2026). Exploiting knowledge graph communities to fine-tune large language models. ESWA, 298, 129816. (www)

  @article{ , abstract = {Since the introduction of GPT-2, Large Language Models (LLMs) have proven to be able to handle various tasks with impressive performance. However, they sometimes generate incorrect output or even hallucinations. To overcome this problem, many researchers have investigated the possibility of integrating external factual knowledge, such as that encoded in Knowledge Graphs (KGs), into LLMs. Although there are many approaches in the existing literature that integrate KGs and LLMs in different ways, few of them use KGs to fine-tune LLMs, and none of them systematically use KG substructures. In this paper, we propose CoFine (Community-Based Fine-Tuner), an approach to fine-tune an LLM using the communities of a KG. CoFine works as follows: it first divides the KG into communities, each of which contains a homogeneous portion of the knowledge expressed by the KG. It then uses these communities to fine-tune the LLM. This way of proceeding allows LLM fine-tuning to focus on specific homogeneous information contained in the KG expressed by each community. CoFine allows the LLM to achieve a very high accuracy in knowledge completion tasks. This is evidenced by comparisons between CoFine and a baseline LLM fine-tuning approach, which showed that our approach achieves better results for all metrics considered with several KG.}, archive = {J_ESWA}, author = {Alessia Amelio and Christopher Buratti and Michele Marchetti and Davide Traini and Domenico Ursino and Luca Virgili}, doi = {10.1016/j.eswa.2025.129816}, journal = {Expert Systems with Applications}, month = {3}, pages = {129816}, shortjournal = {Expert Syst. Appl.}, title = {Exploiting knowledge graph communities to fine-tune large language models}, volume = {298}, year = {2026}, }
• (2026). Multi-scale fusion graph convolutional networks. ESWA, 298, 129815. (www)

  @article{ , abstract = {Graph analysis methods, as important tools for mining complex information, have made remarkable progress driven by graph neural networks (GNNs). However, existing approaches still face challenges in handling complex topological structures and multi-dimensional node features, making it difficult to fully capture deep-level feature and structural information. When analyzing attribute networks, a key challenge is how to effectively integrate node attribute features with graph topological structure information. To address this issue, this paper proposes a multi-scale fusion graph convolutional network (MSF-GCN) method. This method combines shallow and deep convolution strategies while adaptively fusing information across three parallel channels — the original topological structure, a feature-derived graph, and a deep-combination channel that captures shared depth information between them. An autoencoder is employed to reconstruct the adjacency matrix, enhancing the representation capability of the network. Additionally, an attention mechanism is introduced to dynamically assign weights to attribute and structural features at different scales, optimizing node representation. Experimental results demonstrate that, in node classification tasks across multiple benchmark datasets, MSF-GCN achieves outstanding performance, strongly validating the effectiveness and robustness of the proposed method.}, archive = {J_ESWA}, author = {Zhi Kong and Jie Ren and Lifu Wang and Ge Guo}, doi = {10.1016/j.eswa.2025.129815}, journal = {Expert Systems with Applications}, month = {3}, pages = {129815}, shortjournal = {Expert Syst. Appl.}, title = {Multi-scale fusion graph convolutional networks}, volume = {298}, year = {2026}, }
• (2026). Integration path coverage testing of the changing driver code guided by knowledge graph. ESWA, 298, 129814. (www)

  @article{ , abstract = {Device driver code often includes several hardware driver codes. Changes to hardware driver code may thus introduce security risks to device operation. Fortunately, integration target paths are formed by multiple unit functions. These paths reflect the global execution logic of a device’s driver code. Test cases generated for these paths can detect how changing driver code affects the device’s overall operation. Therefore, this paper proposes an approach to integration path coverage testing of the changing driver code guided by knowledge graph. The proposed approach is used to generate test cases covering integration target paths that include changing driver code, queried from the knowledge graph. In the proposed approach, we first statically analyze a device’s hardware driver code, and extract defined entities and their relationships to construct a knowledge graph. From the constructed knowledge graph, we mine entities associated with the changing hardware driver code, which serve as integration target paths. The generation model of test cases is established, which is solved by an intelligent optimization algorithm, thereby generating test cases covering integration target paths. The proposed approach is applied to several open-source driver codes, and compared with several other approaches. Experiments show that the accuracy of both the knowledge graph and the integration target paths has reached 100 %. In addition, the efficiency and effectiveness of test case generation using proposed approach have improved by 19.08 % to 24.54 % and 9 % to 23 %, respectively.}, archive = {J_ESWA}, author = {Baicai Sun and Miao Rong and Gaige Wang and Dunwei Gong}, doi = {10.1016/j.eswa.2025.129814}, journal = {Expert Systems with Applications}, month = {3}, pages = {129814}, shortjournal = {Expert Syst. Appl.}, title = {Integration path coverage testing of the changing driver code guided by knowledge graph}, volume = {298}, year = {2026}, }
• (2026). An efficient deep template matching and in-plane pose estimation method via template-aware dynamic convolution. ESWA, 298, 129813. (www)

  @article{ , abstract = {In industrial inspection and component alignment tasks, template matching requires efficient estimation of a target’s position and geometric state (rotation and scaling) under complex backgrounds to support precise downstream operations. Traditional methods rely on exhaustive enumeration of angles and scales, leading to low efficiency under compound transformations. Meanwhile, most deep learning-based approaches only estimate similarity scores without explicitly modeling geometric pose, making them inadequate for real-world deployment. To overcome these limitations, we propose a lightweight end-to-end framework that reformulates template matching as joint localization and geometric regression, outputting the center coordinates, rotation angle, and independent horizontal and vertical scales. A Template-Aware Dynamic Convolution Module (TDCM) dynamically injects template features at inference to guide generalizable matching. The compact network integrates depthwise separable convolutions and pixel shuffle for efficient matching. To enable geometric-annotation-free training, we introduce a rotation-shear-based augmentation strategy with structure-aware pseudo labels. A lightweight refinement module further improves angle and scale precision via local optimization. Experiments show our 3.07M model achieves high precision and ∼ 14 ms inference under compound transformations. It also demonstrates strong robustness in small-template and multi-object scenarios, making it highly suitable for deployment in real-time industrial applications. The code is available at: https://github.com/ZhouJ6610/PoseMatch-TDCM .}, archive = {J_ESWA}, author = {Ke Jia and Ji Zhou and Hanxin Li and Zhigan Zhou and Haojie Chu and Xiaojie Li}, doi = {10.1016/j.eswa.2025.129813}, journal = {Expert Systems with Applications}, month = {3}, pages = {129813}, shortjournal = {Expert Syst. Appl.}, title = {An efficient deep template matching and in-plane pose estimation method via template-aware dynamic convolution}, volume = {298}, year = {2026}, }
• (2026). GraphShield: Advanced dynamic graph-based malware detection using graph neural networks. ESWA, 298, 129812. (www)

  @article{ , abstract = {The rising complexity of modern malware-such as polymorphic, fileless, and sandbox-aware variants-has severely diminished the reliability of conventional detection techniques. Models based on sequential data frequently miss intricate behavioral patterns and long-range dependencies, resulting in poor accuracy and limited adaptability to new threats. This paper introduces GraphShield, a graph-centric behavioral detection framework that identifies malware with high precision by analyzing dynamic API call sequences. GraphShield converts raw API calls into temporal graphs, applies semantic vectorization, and leverages attention mechanisms to extract both localized activity and extended behavioral correlations, directly addressing the weaknesses of earlier systems. We design and assess multiple Graph Neural Network (GNN) variants, including Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), Graph Isomorphism Networks (GINs), and Transformer-based architectures combining convolutional, recurrent, and autoencoding layers. These models capture structural and temporal traits of execution traces using both classification-only and combined classification-reconstruction strategies. To enhance transparency, we incorporate GNN interpretation tools that isolate key API call subgraphs and critical decision pathways, making detection outcomes explainable for analysts. GraphShield is trained on 300,000 balanced instances and tested on a separate 200,000-sample holdout set, achieving over 58 % improvement in accuracy over advanced sequence-driven deep learning models while maintaining a false positive rate under 1 %. Key features include BERT-based API call grouping for reducing dimensionality and a Markov-inspired graph stabilization method for managing graphs of variable length. Our top models attain a 99.5 % F1-score on the test set. GraphShield aligns recent graph learning techniques with operational cybersecurity needs, delivering accurate detection and clear, interpretable results.}, archive = {J_ESWA}, author = {Eslam Amer and Shaker El-Sappagh and Tamer Abuhamad and Bander Ali Saleh Al-Rimy and Alaa Mohasseb}, doi = {10.1016/j.eswa.2025.129812}, journal = {Expert Systems with Applications}, month = {3}, pages = {129812}, shortjournal = {Expert Syst. Appl.}, title = {GraphShield: Advanced dynamic graph-based malware detection using graph neural networks}, volume = {298}, year = {2026}, }
• (2026). LFRSCNet: Skin defect detection based on lightweight flexible residual separable convolutional network. ESWA, 298, 129811. (www)

  @article{ , abstract = {Aircraft skin is prone to surface damage, such as cracks and dents, during long-term service or manufacturing processes. These defects not only threaten structural integrity but may also pose potential safety hazards. The industrial sector continually explores more efficient and precise detection methods to address this issue. Therefore, this paper proposes a skin defect detection method based on a lightweight and flexible residual separation convolutional network to improve detection accuracy and efficiency. Therefore, this paper proposes a skin defect detection method based on a lightweight flexible residual separable convolution network to improve detection accuracy and efficiency. First, a lightweight flexible residual separable convolution module (LFRCM) is designed, which effectively integrates multi-modal features by combining multi-scale receptive fields with an adaptive channel attention mechanism; at the same time, a lightweight backbone network based on PP-LCNet is constructed, employing a collaborative optimization strategy of depthwise separable convolutions and the h-swish activation function to significantly enhance inference speed while maintaining detection accuracy; finally, the MPDIoU metric criterion is introduced, which effectively improves target localization accuracy by implementing a center point offset penalty mechanism. Experiments on the self-built professional dataset SD-DET and the public dataset GC10-DET show that the model achieves mAP@0.5 of 99.5% and 86.2%, respectively, demonstrating significant advantages over mainstream detection models. Systematic ablation experiments confirm the synergistic effect of various innovative modules. Finally, verification experiments are conducted on the AIRCRAFT skin defect dataset, achieving an mAP@0.5 of 30.7%. Quantitative analysis and comparative experiments verify that LFRSCNet can achieve detection accuracy breakthroughs while maintaining low parameter counts and computational costs. Its balanced accuracy-efficiency characteristics provide an efficient and reliable solution for surface defect detection in industrial scenarios.}, archive = {J_ESWA}, author = {Zhenyu Lu and Jue Wang and Jiteng Zhu and Yuwen Sun}, doi = {10.1016/j.eswa.2025.129811}, journal = {Expert Systems with Applications}, month = {3}, pages = {129811}, shortjournal = {Expert Syst. Appl.}, title = {LFRSCNet: Skin defect detection based on lightweight flexible residual separable convolutional network}, volume = {298}, year = {2026}, }
• (2026). The knowledge-driven adaptive late acceptance iterative hill-climbing heuristics for the bus and ADR collaborative delivery problem. ESWA, 298, 129810. (www)

  @article{ , abstract = {Urban-rural bus transit services encounter a dilemma between the necessity for enhanced services and the challenge of low profitability due to scant travel demand. Combining freight transportation with passenger services can enhance the efficiency and profitability of buses in both urban and rural areas, while also reducing environmental impacts. A case in point is the integration of freight deliveries into rural bus networks in China. Concurrently, with the advancement of autonomous delivery robot (ADR) technology, there is a growing deployment of ADRs for last-mile delivery purposes. In this paper, we have studied a new collaborative passenger and freight transportation problem involving buses and ADRs, namely, the bus and ADR collaborative delivery problem (BACDP). In this scenario, a bus route transports several ADRs, which carry multiple parcels, to distribution regions for door-to-door delivery, each ADR boards a bus to reach the sub-region and then boards another bus to return to the distribution center. We have proposed a mathematical model for BACDP, which can be decomposed into a master problem and a sub-problem. and the condition that the optimal solution to the master problem is also the optimal solution to the original problem has been proved. To tackle the BACDP effectively, we designed a novel three-stage iterative method, guided by adaptive late acceptance hill-climbing heuristics (ALAHH). Specifically, at the first stage, the k-means++ and Hamiltonian graph-guided algorithms are used to cluster customers; at the second stage, the variable neighborhood search plans the ADRs’ routes; at the third stage, we utilize the solver to address subproblems, and the evaluation and invocation mechanism is proposed to achieve the efficient utilization of solvers. Extensive experiments have been conducted on synthetic instances of varying scales to investigate the efficiency of ALAHH. The experimental results demonstrate that the objective values and the computation time are significantly lower than those of SA and LAHC, and our algorithm has achieved the best solutions for 16 problems to date. Additionally, the impacts of two key parameters and mechanisms have been analyzed, and further validation of the robustness of the algorithm parameters and the effectiveness of the mechanisms.}, archive = {J_ESWA}, author = {Lijun Pan and Changshi Liu and Yifan Zhang and Shun Li}, doi = {10.1016/j.eswa.2025.129810}, journal = {Expert Systems with Applications}, month = {3}, pages = {129810}, shortjournal = {Expert Syst. Appl.}, title = {The knowledge-driven adaptive late acceptance iterative hill-climbing heuristics for the bus and ADR collaborative delivery problem}, volume = {298}, year = {2026}, }
• (2026). Future-aware user intent modeling with knowledge distillation for sequential recommendation. ESWA, 298, 129809. (www)

  @article{ , abstract = {With the continuous development of recommender systems, leveraging users’ future interactions to capture the evolution of preferences in sequential behaviors has become a key focus in Sequential Recommendation (SR). Although current approaches have made notable progress, several important challenges persist: 1) An excessive dependence on historical interactions limits the ability to effectively model the evolution of user preferences. 2) There is a lack of mechanisms to capture the dynamic transition path between historical and future behaviors. 3) Existing denoising techniques often fail to fully consider the semantic information in behavior denoising. To address these issues, we propose an SR framework—Future Distillation Recommendation (FDistRec), which introduces the following strategies: To tackle Limitation 1, we incorporate future interaction data to help alleviate the constraints posed by relying on past behaviors. To address Limitation 2, we adopt a distillation framework along with the Causal Mixture-of-Experts (CMoE) module, effectively capturing the evolution and shift path in user interests. To overcome Limitation 3, we present a Masked Denoising Variational Auto Encoder (MDVAE) model that applies semantic-aware masking to perturb user behaviors, facilitating generative preference modeling and noise-aware interest recovery. Experiments on five real-world datasets demonstrate the superior performance of FDistRec compared to the SOTA baselines on SR.}, archive = {J_ESWA}, author = {Xinhua Wang and Xiaodi Liu and Wensheng Sun and GuiyuanJiang and Feng Yuan and Lei Guo}, doi = {10.1016/j.eswa.2025.129809}, journal = {Expert Systems with Applications}, month = {3}, pages = {129809}, shortjournal = {Expert Syst. Appl.}, title = {Future-aware user intent modeling with knowledge distillation for sequential recommendation}, volume = {298}, year = {2026}, }
• (2026). Simultaneous multi-objective path planning with cumulative hazardous dosage constraint for mobile detection robots in complex environments. ESWA, 298, 129808. (www)

  @article{ , abstract = {Path planning for a mobile robot operating in complex hazardous environments often requires simultaneous optimization of multiple objectives, such as minimal path length, low energy consumption, and safe passing under cumulative hazard dosage (CHD) constraint. Available methods, involving deterministic algorithms and meta -heuristic algorithms, have drawbacks in solving this problem to meet these simultaneous requirements. A flexible jump point search strategy (FlexJPS) is proposed to address the problem. In the scheme, jump points are divided into two categories, dominant waypoints (DWPs) and linkage jump points (LJPs). Each DWP is flexibly assigned to one limited zone and all the limited zones are distributed in a spaced form in the grid-based environment. The LJPs are generated in a limited short range by the modified jump point search rule. The DWPs followed by LJPs are evaluated by a designed multi-operator differential evolution algorithm to achieve the optimal solution meeting the safe passing constraint. Various experiments are carried out to validate the performance of the proposed scheme with comparisons to feasible famous counterparts. Statistical results and achieved planning success rates from the experiments under stringent CHD constraint indicate superior performance of the proposed method. It is concluded that the proposed method contributes a high-efficacy multi-objective path planning method for a mobile robot operating in complex environments with CHD constraint.}, archive = {J_ESWA}, author = {Xiankun Lin and Xiaoting Peng and Linsen Liang}, doi = {10.1016/j.eswa.2025.129808}, journal = {Expert Systems with Applications}, month = {3}, pages = {129808}, shortjournal = {Expert Syst. Appl.}, title = {Simultaneous multi-objective path planning with cumulative hazardous dosage constraint for mobile detection robots in complex environments}, volume = {298}, year = {2026}, }
• (2026). A multi-domain feature-based epileptic seizure prediction method using EEG source estimation and graph theory. ESWA, 298, 129807. (www)

  @article{ , abstract = {Epilepsy is a disorder caused by abnormal discharges of cerebral neurons, affecting 50 million people worldwide. Most existing research on seizure prediction remains at the scalp level. To explore and harness the potential of information flow among cortical regions as well as the intrinsic brain networks and functional mechanisms for seizure prediction, this study proposes a novel multi-domain feature fusion seizure prediction framework based on Electroencephalography (EEG) source estimation and graph theory. Specifically, dSPM source estimation and singular value decomposition (SVD) are first applied to extract 44 subcortical regions defined by the “HCPMMP1_combined” atlas. Brain networks are then constructed using coherence (COH) and phase lag index (PLI), from which specific network topological features based on graph theory are calculated. These features are further extended to higher-order brain networks to enhance connectivity modeling. We also build a multi-domain feature hybrid (MFH) prediction model that adopts a multi-branch structure. One branch employs hypergraph convolution attention along with time–frequency node features derived from continuous wavelet transform (CWT) to capture high-order spatial correlations; another branch inputs the temporal signals from cortical brain regions and combines the hybrid Mamba2-Transformer expert (HMT) model with a frequency-domain dot-product channel attention network (FDCA). By fusing these branches, the model demonstrates satisfactory results across multiple age-groups in the epilepsy EEG dataset of Gansu Province Central Hospital, as well as in the CHB − MIT dataset. This framework highlights the potential of integrating source estimation, hypergraph analysis, and multi-domain feature learning for personalized seizure prediction.}, archive = {J_ESWA}, author = {Bingyang Ji and Wenwen Chang and Guanghui Yan and Dandan Li and Rong Yin and Xuan Liu and Yaxuan Wei}, doi = {10.1016/j.eswa.2025.129807}, journal = {Expert Systems with Applications}, month = {3}, pages = {129807}, shortjournal = {Expert Syst. Appl.}, title = {A multi-domain feature-based epileptic seizure prediction method using EEG source estimation and graph theory}, volume = {298}, year = {2026}, }
• (2026). Multi-hop commonsense knowledge injection framework for zero-shot commonsense question answering. ESWA, 298, 129806. (www)

  @article{ , abstract = {Zero-shot commonsense question answering (QA) task is to evaluate the general reasoning ability of the language model without training on the specific datasets. The existing zero-shot framework transforms triples within the commonsense knowledge graphs (KGs) into QA-format samples, serving as a pre-training data source to integrate commonsense knowledge into the language model. However, this approach still faces the following challenges: 1) The model trained from synthetic QA generated from triples lacks the multi-hop commonsense knowledge required for handling complex QA problems. 2) Ambiguity caused by confusing commonsense knowledge within synthetic QA, making it challenging for models to discern semantically similar entities. To address the above problem, we propose a novel M ulti-hop C ommonsense K nowledge I njection Framework (MCKI). Specifically, we draw inspiration from human complex reasoning thinking and further propose a synthetic multi-hop commonsense QA generation method. Meanwhile, we introduce negative samples with high confusion in synthetic QA, and then use contrastive learning to improve the model’s ability to distinguish similar commonsense knowledge. Extensive experiments on five commonsense question answering benchmarks demonstrate that our framework achieves state-of-the-art performance, surpassing existing methods, including large language models like GPT3.5 and ChatGPT.}, archive = {J_ESWA}, author = {Xin Guan and Jiuxin Cao and Biwei Cao and Qingqing Gao and Bo Liu}, doi = {10.1016/j.eswa.2025.129806}, journal = {Expert Systems with Applications}, month = {3}, pages = {129806}, shortjournal = {Expert Syst. Appl.}, title = {Multi-hop commonsense knowledge injection framework for zero-shot commonsense question answering}, volume = {298}, year = {2026}, }
• (2026). CR-GAC: Cross-modal recombination via graph-attention collaborative optimization for multimodal sentiment analysis. ESWA, 298, 129805. (www)

  @article{ , abstract = {Multimodal sentiment analysis integrates linguistic, audio, and visual modalities for predicting human emotional states. However, current algorithms encounter three challenges: limitations in adjacency matrix modeling, noise interference and modality imbalances in cross-modal attention, and inefficient cross-modal feature alignment. To address these, we propose the C ross-modal R ecombination via G raph- A ttention C ollaborative Optimization (CR-GAC) by unifying graph and sequence learning in a collaborative framework. Specifically, we first design the modality-adaptive M ultimodal G raph C onstruction (MGC) to tackle the first challenge. For the linguistic modality, a local sparse graph based on a K-Nearest Neighbors-Radial Basis Function kernel is designed to preserve fine-grained semantics; for the audio and visual modalities, a low-rank representation method combined with nuclear norm regularization is designed to capture latent cross-sample structures via singular value decomposition, while suppressing noise interference. Modalities that have been processed are then input into graph attention networks to achieve higher-order feature aggregation. Next, we construct the L anguage-guided H ierarchical C ross-modal I nteraction (LHCI) to tackle the second challenge, which leverages bidirectional cross-modal attention and multi-level Transformer blocks to hierarchically enhance feature representations. Subsequently, the H igh-level M ultimodal F eature C ontainer (HMFC) iteratively accumulates multi-grained semantics, providing a high-level feature pool for fusion. Finally, the dynamic matching-based H igh-level F eature R ecombination (HFR) is designed to tackle the third challenge, which uses the linguistic feature as an anchor to achieve semantically controllable explicit alignment and flexible implicit alignment by matching the most relevant features. Experimental results show our model achieves state-of-the-art performance on CMU-MOSI and CMU-MOSEI datasets, and demonstrates generalization capability on CH-SIMS dataset.}, archive = {J_ESWA}, author = {Haoran Chen and Jiapeng Liu and Zuhe Li and Yushan Pan and Hongwei Tao and Huaiguang Wu and Yunyang Wang and Chenguang Yang}, doi = {10.1016/j.eswa.2025.129805}, journal = {Expert Systems with Applications}, month = {3}, pages = {129805}, shortjournal = {Expert Syst. Appl.}, title = {CR-GAC: Cross-modal recombination via graph-attention collaborative optimization for multimodal sentiment analysis}, volume = {298}, year = {2026}, }
• (2026). Sustainable time-dependent intermodal hub-and-spoke logistic network considering hub failure: A mathematical model and a hybrid artificial bee colony algorithm. ESWA, 298, 129804. (www)

  @article{ , abstract = {This paper addresses the design of a sustainable hub-and-spoke logistics network that integrates intermodal transportation between the hubs, hub failures, time dependency, and environmental parameters. Accordingly, we propose a novel mixed-integer linear programming (MILP) model and a hybrid artificial bee colony-based algorithm (HABCb) to minimize transportation costs and emissions in robust network configurations. The model is the first to simultaneously integrate intermodality, sustainability metrics, and hub disruption scenarios within a single framework. Computational experiments using real-life data from Turkey demonstrate that the proposed HABCb approach outperforms both genetic algorithm (GA) and artificial bee colony (ABC) algorithm. On medium-sized problem sets, it achieves average cost reductions of 7% compared to GA and 10% compared to ABC algorithm, while on large-sized problems the reductions are 10% and 15%, respectively. Furthermore, the HABCb approach provides faster convergence and higher-quality solutions for larger problem sizes. The findings highlight the practical and theoretical insights of incorporating sustainability, intermodality, and robustness into hub-and-spoke network design.}, archive = {J_ESWA}, author = {Burcu Tokbay Erkek and Salih Himmetoğlu and Yılmaz Delice and Emel Kızılkaya Aydoğan}, doi = {10.1016/j.eswa.2025.129804}, journal = {Expert Systems with Applications}, month = {3}, pages = {129804}, shortjournal = {Expert Syst. Appl.}, title = {Sustainable time-dependent intermodal hub-and-spoke logistic network considering hub failure: A mathematical model and a hybrid artificial bee colony algorithm}, volume = {298}, year = {2026}, }
• (2026). A belief rule-based system for online and centralized collaborative performance assessment of networked physical systems subject to nonideal channels. ESWA, 298, 129803. (www)

  @article{ , abstract = {Networked physical systems (NPSs) are widely applied in modern engineering practices characterized by intensive domain knowledge and imperfect observational data. Meanwhile, collaborative performance assessment provides strong support for them to operate safely and stably over a long period of time. For a specific NPS and its corresponding online and centralized collaborative performance assessment system, the existence of interference and noise in the real-world channel that is rather nonideal inevitably obstructs the smooth progress of the assessment. To this end, in this paper, a symbolic systematic solution is proposed resorting to an improved version of the belief rule base with continuous inputs (BRB-CI). First, the extrapolation module is enhanced by integrating a matched filtering-based link. Second, the existing robustness analysis for systems based on the fundamental belief rule base is extended to systems based on the BRB-CI. Third, the optimization module is ameliorated by designing a multimetric-balanced pattern of the grey wolf optimizer with interpretability reinforcement. Ultimately, by choosing an instance of NPSs in the field of aerospace with continuous time dynamics, pertinent empirical studies are carried out to substantiate the good engineering practicability of our proposal. Note that this paper is the first piece inquiring into belief rule-based systems such a class of expert systems for online and centralized cooperative performance assessment of NPSs with continuous time dynamics such an application, with considerable attention paid to the nonideality of real-world channels.}, archive = {J_ESWA}, author = {Haoran Zhang and Lining Xing and Jian Wu and Ruohan Yang and Zhichao Feng}, doi = {10.1016/j.eswa.2025.129803}, journal = {Expert Systems with Applications}, month = {3}, pages = {129803}, shortjournal = {Expert Syst. Appl.}, title = {A belief rule-based system for online and centralized collaborative performance assessment of networked physical systems subject to nonideal channels}, volume = {298}, year = {2026}, }
• (2026). E2D-GS: Event-enhanced deblurring gaussian splatting. ESWA, 298, 129802. (www)

  @article{ , abstract = {In recent years, implicit neural representations and explicit 3D Gaussian Splatting(3DGS) have demonstrated substantial advancements in the domain of novel view synthesis. Nevertheless, the efficacy of these approaches is predominantly contingent upon the availability of well-defined, clear imagery and precise camera pose information. Consequently, they exhibit a pronounced susceptibility to motion blur, which impedes the rendering of sharp images. Event cameras, which measure intensity changes with microsecond temporal precision, possess an inherent robustness to motion-induced blur. This characteristic offers new avenues for 3D reconstruction in challenging scenarios characterized by high-speed motion or low-light conditions. This paper introduces E2D-GS, a novel algorithm for deblurring and reconstruction based on event cameras and 3D Gaussian Splatting. To enhance reconstruction accuracy, our proposed framework leverages event streams to physically model the formation process of motion blur. This is achieved by optimizing the discrepancy between synthesized data and the observed blurry images, while simultaneously recovering the camera’s motion trajectory. Additionally, to enhance robustness in real-world scenarios, this paper proposes a differential consistency module. This module effectively mitigates noise within the event data and regularizes the optimization of Gaussian parameters, thereby improving reconstruction quality under non-ideal conditions. Comprehensive experimental evaluations on both simulated and real-world benchmarks validate the proposed method’s capability to reconstruct latent sharp imagery via the learned 3DGS representations, and further demonstrate its capacity for stable reconstruction under adverse scenarios. The results show that our approach surpasses the performance of previous works.}, archive = {J_ESWA}, author = {Lifeng Lin and Shuangjie Yuan and Lu Yang}, doi = {10.1016/j.eswa.2025.129802}, journal = {Expert Systems with Applications}, month = {3}, pages = {129802}, shortjournal = {Expert Syst. Appl.}, title = {E2D-GS: Event-enhanced deblurring gaussian splatting}, volume = {298}, year = {2026}, }
• (2026). Adaptive gated universal information extraction for chinese legal texts. ESWA, 298, 129801. (www)

  @article{ , abstract = {Constructing knowledge graphs in legal domains requires simultaneous extraction of entities and relations. To reduce repeated modeling in traditional approaches, we adopt the Universal Information Extraction (UIE) model as a foundation and propose an enhanced variant named Adaptive Gated Universal Information Extraction (AGUIE). This study develops a new decoder based on the Adaptive Focusing Gated Attention Unit (AFGAU). This unit enhances the standard Gated Attention Unit (GAU) by integrating two key components—learnable dynamic convolution and reset/update gating mechanisms. Moreover, the study employs a cross-pointer structure as the output layer to better identify information boundaries. To support this study, we construct a domain specific dataset for extracting key information from legal judgment documents. Systematic comparative analysis and ablation studies demonstrate that AGUIE achieves significant performance gains over baseline UIE, with an F1 score of 85.56% on our legal judgment documents dataset. Additionally, we evaluate the model’s generalization on public datasets such as ACE04, ACE05, and CoNLL04, covering both entity recognition and relation extraction tasks. Experimental results indicate that AGUIE demonstrates competitive results with recent studies on ACE04-Ent and CoNLL04, outperforms them on the ACE05 dataset, achieving F1 scores of 87.19% on ACE05-Ent and 79.29% on ACE05-Rel. In conclusion, AGUIE is a reliable and effective solution for universal information extraction in both legal and general domains.}, archive = {J_ESWA}, author = {Yabo Liu and Yatong Zhou and Kuo-Ping Lin}, doi = {10.1016/j.eswa.2025.129801}, journal = {Expert Systems with Applications}, month = {3}, pages = {129801}, shortjournal = {Expert Syst. Appl.}, title = {Adaptive gated universal information extraction for chinese legal texts}, volume = {298}, year = {2026}, }
• (2026). Radiomics knowledge-driven deep learning framework for breast cancer ultrasound video diagnosis. ESWA, 298, 129800. (www)

  @article{ , abstract = {Early and accurate diagnosis of breast cancer is essential. However, the absence of expert knowledge guidance often hampers the performance and clinical applicability of existing automated cancer diagnostic models. Radiomics indicators serve as important references for physicians during clinical diagnosis; nevertheless, limited research has incorporated such domain knowledge into deep learning frameworks. Therefore, we propose a radiomics knowledge-driven deep learning framework for breast cancer diagnosis based on dual-amplitude contrast-enhanced ultrasound (CEUS) videos. The framework employs a three-branch video classification network augmented by a three-step knowledge guidance process. This process begins with radiomics knowledge extraction, deriving both numerical and categorical features from the video data. Quantitative knowledge fusion then integrates these numerical features using numerical bin encoding and the spatio-temporal causal module. Finally, qualitative knowledge fusion incorporates categorical features via label embedding and the multi-label calibrated loss, aligning model outputs with categorical features. Extensive experiments on triple-class cancer classification demonstrate superior performance, achieving an AUC of 94.50 % ± 1.50 % and an accuracy of 83.88 % ± 1.94 %, while providing symptom analysis results as clinical decision support. Furthermore, our framework achieves consistent performance gains when applied to other deep learning models, highlighting its broad applicability and potential for advancing medical diagnosis. Code is available at https://github.com/tobenan/radiomics .}, archive = {J_ESWA}, author = {Dinghao Guo and Zheng Xue and Dali Chen and Chunyu Lu and Jun Fu and Jizhong Yuan and Ying Huang}, doi = {10.1016/j.eswa.2025.129800}, journal = {Expert Systems with Applications}, month = {3}, pages = {129800}, shortjournal = {Expert Syst. Appl.}, title = {Radiomics knowledge-driven deep learning framework for breast cancer ultrasound video diagnosis}, volume = {298}, year = {2026}, }
• (2026). A hybrid graph and LLM approach for measuring scientific novelty via knowledge recombination and propagation. ESWA, 298, 129794. (www)

  @article{ , abstract = {Scientific novelty constitutes a fundamental catalyst for both disciplinary innovation and interdisciplinary progress. Nevertheless, prevailing approaches to novelty assessment predominantly emphasize a single analytical dimension–either the semantic content of the focal paper or its cited references. Content-based methodologies frequently fail to incorporate the foundational knowledge cited by the target publication, whereas reference-based strategies tend to disregard the intrinsic conceptual contributions of the focal work itself. To address this limitation, the present study introduces a hybrid graph and large language model approach to jointly capture and integrate knowledge embedded in both the focal paper and its cited literature. The proposed method, which integrates knowledge recombination and propagation, is structured into four primary stages. First, prompt-based extraction techniques using general LLMs are applied to extract knowledge. Second, a Reference Knowledge Combination Network (RKCN) is constructed to model the knowledge referenced by the focal paper. Third, the RKCN is initialized with representations generated by SciDeBERTa(CS), and a graph attention network is employed to propagate knowledge across the network. Finally, the novelty of the focal paper is quantified by aggregating the novelty scores of all internal knowledge combinations based on the propagated representations. Experimental evaluation in the domain of artificial intelligence (AI) demonstrates that the proposed method significantly outperforms existing baseline approaches in quantifying scientific novelty. Additional ablation studies further validate the contribution of the knowledge propagation module. A case study illustrates the interpretability of our approach, and a cross-field validation in Biomedical Engineering (BME) domain highlights its robustness and cross-domain generalizability. A multi-dimensional comparative analysis between award-winning and non-award papers further reveals that the former generally incorporate a larger volume of knowledge and exhibit greater diversity in knowledge combinations. Moreover, while both groups encompass knowledge combinations spanning a wide range of novelty, award-winning papers display a stronger concentration at higher novelty levels, in contrast to the more uniform distribution observed in non-award papers. Data, code, and more detailed results are publicly available at: https://github.com/haihua0913/graphLLM4ScientificNovelty .}, archive = {J_ESWA}, author = {Zhongyi Wang and Zeren Wang and Guangzhao Zhang and Jiangping Chen and Markus Luczak-Roesch and Haihua Chen}, doi = {10.1016/j.eswa.2025.129794}, journal = {Expert Systems with Applications}, month = {3}, pages = {129794}, shortjournal = {Expert Syst. Appl.}, title = {A hybrid graph and LLM approach for measuring scientific novelty via knowledge recombination and propagation}, volume = {298}, year = {2026}, }
• (2026). Graph-based approaches for rumor detection in social networks: A systematic review. ESWA, 298, 129786. (www)

  @article{ , abstract = {Increased public anxiety and fear, disrupted decision-making, social instability, and other significant societal challenges are the results of the rapid spread of rumors on social media platforms. The unique characteristics of these platforms contribute to the rapid spread of both verified and unverified information. These pressing issues highlight the need to develop advanced technologies for early detection and prevention of rumors. This paper presents a systematic review of graph-based approaches for rumor detection in social networks, analyzing 53 studies published between 2018 and 2025. The selected studies are comprehensively reviewed with a focus on graph models and the integration of propagation structure, social, temporal, and content features, which enhances detection accuracy. This review critically evaluates the effectiveness of various methods, highlighting their strengths, limitations, and key challenges. The key contributions of this paper include: (i) an in-depth analysis of current graph-based rumor detection approaches (ii) a categorization of graph models and feature extraction strategies, (iii) the identification of major challenges and research gaps, and (iv) recommendations for future research to develop scalable, robust, and accurate early rumor detection systems. The findings of this study provide valuable insights for researchers aiming to advance the state-of-the-art in fighting misinformation on social networks.}, archive = {J_ESWA}, author = {Fatima Al-Thulaia and Seyyed Alireza Hashemi Golpayegani}, doi = {10.1016/j.eswa.2025.129786}, journal = {Expert Systems with Applications}, month = {3}, pages = {129786}, shortjournal = {Expert Syst. Appl.}, title = {Graph-based approaches for rumor detection in social networks: A systematic review}, volume = {298}, year = {2026}, }
• (2026). Spatiotemporal online fuzzy modeling with knowledge-driven differential evolution automatic clustering for distributed parameter systems. ESWA, 298, 129785. (www)

  @article{ , abstract = {Distributed parameter systems are prevalent in various industrial processes and attract significant attention. However, these systems exhibit complex spatiotemporal coupling characteristics, and effectively determining the fuzzy rules of the antecedent set is crucial for improving modeling performance. Traditional clustering methods typically rely on empirical heuristics and are unable to adapt to dynamic system characteristics under changing environments. In high-dimensional and nonlinear scenarios, the number of fuzzy rule combinations grows exponentially, significantly increasing computational complexity. Therefore, an online spatiotemporal three-dimensional fuzzy modeling method based on knowledge-driven differential evolution automatic clustering and extreme learning machine (3D-OSADE-ELM) is proposed for the complex nonlinear distributed parameter system. First, an automatic clustering mechanism based on differential evolution and extreme learning machine initializes the fuzzy rules within the three-dimensional fuzzy system. Subsequently, a knowledge-driven archiving mechanism dynamically updates the fuzzy rules of the antecedent set during the online incremental learning phase. Finally, the spatial basis function is obtained by learning the output weight of the online extreme learning machine. The validation experiments conducted on the rapid thermal chemical vapor deposition reactor system and the nonisothermal packed-bed system demonstrate the effectiveness and superiority of the proposed method.}, archive = {J_ESWA}, author = {Gang Zhou and Xianxia Zhang and Bing Wang}, doi = {10.1016/j.eswa.2025.129785}, journal = {Expert Systems with Applications}, month = {3}, pages = {129785}, shortjournal = {Expert Syst. Appl.}, title = {Spatiotemporal online fuzzy modeling with knowledge-driven differential evolution automatic clustering for distributed parameter systems}, volume = {298}, year = {2026}, }
• (2026). CG-TRAN: A novel multi-label retinal disease classification model with partially known pathologies. ESWA, 298, 129784. (www)

  @article{ , abstract = {Early detection of retinal diseases is vital to preventing partial or permanent blindness. However, the diagnostic process is often impeded by the complexity of interrelated lesions and the challenge of incomplete or missing pathology labels, which require specialized expertise in ophthalmic diagnosis. To address these limitations, we propose CG-Tran, a novel multi-label classification model that leverages partially known pathology information to diagnose retinal diseases. This approach integrates a pathology graph neural network with graph-based feature extraction to handle partially known pathologies, enabling more accurate multi-label classification of retinal diseases. To model the intricate interrelationships among ocular diseases, CG-Tran employs BERT-GNN to learn label interactions and construct a comprehensive fundus pathology graph. Additionally, an enhanced attention mechanism incorporates known pathology label features, bridging the gap between incomplete pathology information and fundus image data. These innovations collectively empower the model to overcome the challenges of missing or incomplete pathology labels. The model’s performance is rigorously evaluated on the Multilabel Retinal Disease (MuReD) dataset. Results demonstrate that CG-Tran significantly improves diagnostic accuracy, especially as more pathology labels become available. Under conditions with 0% and 75% partially known labels, CG-Tran achieves mean average precision (mAP) scores of 69.9% and 72.1%, respectively—outperforming the baseline model by 1.0% and 1.9%. This innovative architecture excels in multi-label classification tasks, particularly in recognizing and distinguishing complex and interrelated retinal lesions with partially known pathology. It offers a promising solution for early detection and accurate diagnosis of retinal diseases, addressing critical limitations in existing diagnostic methods.}, archive = {J_ESWA}, author = {Jia Sheng Yang and Zihao Ning and Xu Xiao and Rui Zhong and Chenbo Xia and Ya Ding}, doi = {10.1016/j.eswa.2025.129784}, journal = {Expert Systems with Applications}, month = {3}, pages = {129784}, shortjournal = {Expert Syst. Appl.}, title = {CG-TRAN: A novel multi-label retinal disease classification model with partially known pathologies}, volume = {298}, year = {2026}, }
• (2026). Time-balanced MSE for machinery imbalanced degradation trend prediction. ESWA, 298, 129783. (www)

  @article{ , abstract = {Accurate degradation trend prediction (DTP) is crucial for optimizing equipment operation and maintenance. With the rapid development of artificial intelligence, many data-driven methods have been applied to machinery degradation trend prediction. In practice, most machineries are in the early stages of degradation, with only a few reaching the final stages, leading to a temporal imbalanced data distribution. Current research on imbalanced distributions mainly focuses on classification tasks. However, DTP involves multiple time-dependent continuous targets, making classification-based methods unsuitable. To address this issue, the degradation trend prediction task is reformulated as a multi-task problem and a novel time-balanced Mean Square Error (TBMSE) loss function is proposed. In each prediction task, the Gaussian Mixture Model (GMM) is used to fit the training label distribution. Additionally, the cumulative information noise for each prediction task is modeled using GMM, and an end-to-end network structure is designed to learn the GMM parameters. Experiments are conducted on the IMS bearing dataset and the turboprop engine dataset, demonstrating that the TBMSE loss effectively mitigates the issue of temporal imbalanced distribution in degradation trend prediction.}, archive = {J_ESWA}, author = {Yu-Qiang Wang and Yong-Ping Zhao and Tian-Ding Zhang and Yu-Wei Wang}, doi = {10.1016/j.eswa.2025.129783}, journal = {Expert Systems with Applications}, month = {3}, pages = {129783}, shortjournal = {Expert Syst. Appl.}, title = {Time-balanced MSE for machinery imbalanced degradation trend prediction}, volume = {298}, year = {2026}, }
• (2026). Differentiable histogram-guided unsupervised retinex enhancement for paired low-light images. ESWA, 298, 129782. (www)

  @article{ , abstract = {Most existing low-light image enhancement (LIE) methods rely on expensive paired low-light and normal-light datasets, while unsupervised approaches depend on handcrafted priors to design networks or select similar normal-light images as pseudo-references, limiting their generalization and robustness. To address these challenges, we propose a novel differentiable histogram-guided unsupervised Retinex enhancement (DHURE) method, which leverages the distribution of illumination histograms in real-world scenarios to achieve high-fidelity color preservation and refined brightness distribution across diverse extremely low-light images. DHURE avoids reliance on scene-specific features and effectively captures both fine-grained details and overall brightness information. Specifically, our method consists of two key components: 1) The lightweight architecture of DHURE is composed of Retinex decomposition and illumination enhancement. We perform Retinex decomposition on paired low-light images (PRD) and design the Illumination Histogram-guided Enhancement (IHE) module. Both modules employ lightweight architectures. 2) To fully exploit the adaptive priors inherent in paired low-light images, we introduce a self-supervised reflectance map loss that aligns with the Retinex basis loss. Based on the illumination distribution of real-world normal-light images, we define two unsupervised illumination histogram losses, enabling more generalized and robust enhancement. Extensive and diverse experiments demonstrate that our method achieves competitive performance compared to existing unsupervised LIE approaches, showing superior results on most evaluation metrics. The source code is available at https://github.com/yoonyin/DHURE-main .}, archive = {J_ESWA}, author = {Liyuan Yin and Pingping Liu and Tongshun Zhang and Hongwei Zhao and Qiuzhan Zhou}, doi = {10.1016/j.eswa.2025.129782}, journal = {Expert Systems with Applications}, month = {3}, pages = {129782}, shortjournal = {Expert Syst. Appl.}, title = {Differentiable histogram-guided unsupervised retinex enhancement for paired low-light images}, volume = {298}, year = {2026}, }
• (2026). SUNRISE: Multi-agent reinforcement learning via neighbors’ observations under fully noisy environments. ESWA, 298, 129781. (www)

  @article{ , abstract = {Multi-agent reinforcement learning (MARL) methodologies have achieved notable advancements across diverse domains. Despite these successes, the susceptibility of neural networks to perturbed data and the ubiquity of external attacks in real-world settings, such as sensor noise, pose challenges for MARL approaches. The pivotal issue revolves around the effective transfer of policies learned in idealized simulation environments to the complexities inherent in real-world scenarios. More precisely, when agents are unable to obtain any accurate observations of the external environment throughout the entire policy learning process, the MARL methods cannot learn effective policies. In addressing this issue, we propose a methodology wherein noisy observations from neighboring agents are utilized, with an agent’s own noisy observations serving as surrogate ground truth. This approach facilitates the learning of effective policies by MARL methods in environments characterized by pervasive noise. We design a denoising representation network to filter out the principal state information from environment data characterized by noise to mitigate the adverse effects of noise on the process of policy learning. Then, we integrate the denoising representation network with classic MARL methodologies to learn effective policies within environments characterized by pervasive noise. A series of exhaustive experimental results demonstrate the efficacy of our approach in attenuating the impact of external attacks on the optimization parameters of neural networks during the policy-learning process. Moreover, our methodology exhibits compatibility with classic MARL methods, allowing for the learning of effective policies.}, archive = {J_ESWA}, author = {Kaiyu Wang and Bohao Qu and Menglin Zhang and Xianchang Wang and Ximing Li}, doi = {10.1016/j.eswa.2025.129781}, journal = {Expert Systems with Applications}, month = {3}, pages = {129781}, shortjournal = {Expert Syst. Appl.}, title = {SUNRISE: Multi-agent reinforcement learning via neighbors’ observations under fully noisy environments}, volume = {298}, year = {2026}, }
• (2026). RSUTrajRec: Multi-granularity trajectory recovery based on roadside units sensing. ESWA, 298, 129780. (www)

  @article{ , abstract = {Vehicle mobility trajectories, especially fine-grained trajectories, provide valuable insights for understanding urban dynamics and play a crucial role in intelligent transportation systems and urban planning. Obtaining fine-grained vehicle trajectories can be realized by trajectory recovery, but traditional efforts suffer from defects such as poor privacy protection and low recovery accuracy. To address these issues, we propose a new scenario of trajectory recovery based on roadside unit (RSU) sensing. However, this scenario introduces a significant challenge: recovering high-precision trajectories from the incomplete and unevenly distributed sensing data. To tackle this, we design RSUTrajRec , a multi-granularity trajectory recovery framework that comprises a graph neural network-based module for road information prediction, a Transformer-based module for multi-granularity recovery, and an RSU deployment planning module. Extensive real-world dataset evaluations reveal that RSUTrajRec has a significant advantage in recovering missing vehicle trajectories outside the RSU coverage area. In addition, evaluations also verify that the performance of the trajectory recovery task can be effectively improved by optimizing the RSU deployment plan.}, archive = {J_ESWA}, author = {Xianjing Wu and Xutao Chu and Jianyu Wang and Shengjie Zhao}, doi = {10.1016/j.eswa.2025.129780}, journal = {Expert Systems with Applications}, month = {3}, pages = {129780}, shortjournal = {Expert Syst. Appl.}, title = {RSUTrajRec: Multi-granularity trajectory recovery based on roadside units sensing}, volume = {298}, year = {2026}, }
• (2026). Technology foresight in china’s industrial robotics with MLWS-TF: A machine learning and weak signal-based system. ESWA, 298, 129779. (www)

  @article{ , abstract = {Technology foresight analyses technological trends and potential impacts to provide strategic guidance. However, existing methods either rely on experts to discover emerging directions leading to subjective bias or adopt machine learning to predict without explanation. We propose a Machine Learning and Weak Signal-based Technology Forecasting System (MLWS-TF), which is entirely data-driven to enhance the objectivity of technology foresight and can interpret emerging directions through weak signals. The system adopts a two-phase machine learning model (2P-ML), the first phase identifies papers related to the robotics field, while the second further classifies them into fine-grained research directions. Keywords are extracted from the papers using a Word2Vec-based approach, and a three-dimensional signal classification method (DVI) is developed to quantify the foresight value of keywords across the Diffusion, Visibility, and Impact dimensions, identifying weak signals for technology forecasting. Experiments evaluate various machine learning algorithms, and XGBoost outperforms in constructing the 2P-ML classifier. The model achieved over 90% accuracy, demonstrating its effectiveness in identifying the theme of scientific documents based on textual features. For each research theme, the DVI provides a more comprehensive assessment of signal strength to detect weak signals. Finally, MLWS-TF analyses the growth potential of themes and successfully identifies critical development directions. Our approach offers a novel automated technology foresight system, which completely avoids the subjectivity and dependence on expert judgment that characterize traditional technology foresight approaches, and extends weak signal theory by introducing the Impact dimension to evaluate signal strength.}, archive = {J_ESWA}, author = {Ruihan Wang and Yuhao Zhu}, doi = {10.1016/j.eswa.2025.129779}, journal = {Expert Systems with Applications}, month = {3}, pages = {129779}, shortjournal = {Expert Syst. Appl.}, title = {Technology foresight in china’s industrial robotics with MLWS-TF: A machine learning and weak signal-based system}, volume = {298}, year = {2026}, }
• (2026). Generating realistic pruning solutions for automated grape vine pruning using graph neural networks. ESWA, 298, 129778. (www)

  @article{ , abstract = {In our prior work we showed that graph neural networks (GNNs) can be trained to generate pruning solutions that could direct robotic pruning robots to perform automated cane pruning of wine grape vines. That study introduced the feasibility of the technology but also showed that there were many open questions and issues with the research results that needed to be addressed. In this study we address some of these questions. For example, we answer the question of how would a model like this perform on real vine architectures compared with pruning solutions from real experienced pruners. Our most notable contributions include moving away from a per-cane classification model that attempts to define a single perfect pruning solution, to a model that ranks multiple good solutions and picks the best one. We addressed a key limitation of the previous training data by moving away from synthetic vine architectures to realistic ones recorded from real vines and using pruning solutions collected by expert pruners as our ground-truth. Our primary goal was to show that learning by example using a GNN-based model was a viable approach to automated pruning, even when compared with experienced pruners. We showed robust performance from our model by training on a dataset of 90 pruning solutions generated by expert pruners in the 2022 season, and testing our performance on 117 pruning solutions from an independent set of pruners from the 2021 season. The model was able to correctly score all the pruning solutions from the 2021 dataset as good to very good and none of the expert solutions were classified as poor .}, archive = {J_ESWA}, author = {Jaco Fourie and Jeffrey Hsiao and Oliver Batchelor and Kevin Langbroek and Henry Williams and Richard Green and Armin Werner}, doi = {10.1016/j.eswa.2025.129778}, journal = {Expert Systems with Applications}, month = {3}, pages = {129778}, shortjournal = {Expert Syst. Appl.}, title = {Generating realistic pruning solutions for automated grape vine pruning using graph neural networks}, volume = {298}, year = {2026}, }
• (2026). Disruption-responsive berth allocation and quay crane scheduling with inter-terminal collaboration. ESWA, 298, 129776. (www)

  @article{ , abstract = {Container terminal operations frequently encounter disruptions, including delays, extended handling times, and unscheduled vessel arrivals, all of which necessitate intelligent rescheduling strategies to maintain operational efficiency. This study investigates the integrated problem of disruption-responsive berth allocation and quay crane (QC) scheduling, explicitly considering vessel gathering status and incorporating inter-terminal shifting (ITS) and reassignment to terminals different from its originally designated one (RT) as adaptive response strategies to mitigate these disruptions. A rescheduling model is developed to minimize associated costs. To efficiently solve large-scale problems, an adaptive large neighborhood search (ALNS)-based heuristic is proposed. The effectiveness of the proposed scheme is validated through comparative experiments involving three alternative schemes, highlighting its superior performance. Furthermore, algorithm comparison experiments are conducted to verify the robustness of parameter settings. Computational results demonstrate that the proposed model and algorithm achieve high efficiency and solution quality. Additionally, sensitivity analysis reveals that neglecting vessel gathering status leads to substantial cost increases, particularly in large-scale operations. The integration of ITS and RT proves to be an effective strategy for mitigating disruptions, enhancing scheduling flexibility, and improving operational performance.}, archive = {J_ESWA}, author = {Hongxing Zheng and Zhaoyang Wang and Lingxiao Wu}, doi = {10.1016/j.eswa.2025.129776}, journal = {Expert Systems with Applications}, month = {3}, pages = {129776}, shortjournal = {Expert Syst. Appl.}, title = {Disruption-responsive berth allocation and quay crane scheduling with inter-terminal collaboration}, volume = {298}, year = {2026}, }
• (2026). Mining spatiotemporal dominant co-location patterns. ESWA, 298, 129775. (www)

  @article{ , abstract = {Spatial co-location pattern mining is an important branch of spatial data mining, which can identify spatial features that prevalently occur in proximity. Based on spatial co-location patterns, the research of dominant relationships mining within co-location patterns further considers the influence relationship among features. However, relying solely on spatial data to analyze the positions and distribution of features for mining dominant relationships is insufficient and may lead to incorrect patterns. To address this limitation, this paper introduces the temporal factor into the research of dominant relationships mining and proposes the spatiotemporal dominant co-location pattern mining (STDCPM). At first, we define the concepts of spatiotemporal dominant relationship from both temporal and spatial dimensions, and then propose the spatiotemporal dominant participation index to assess the prevalence of spatiotemporal dominant co-location patterns. Furthermore, we design two algorithms, the spatiotemporal dominant co-location pattern mining algorithm with level-by-level search and its improved version, i.e., the spatiotemporal dominant co-location pattern mining approach based on dual pruning and refining set (STDCPM-DPR), to ensure efficient mining in spatiotemporal datasets. The time complexity, correctness, and completeness of proposed algorithms are discussed. Extensive experiments on real-world datasets demonstrate the effectiveness of STDCPM and the efficiency of STDCPM-DPR algorithm.}, archive = {J_ESWA}, author = {Jiangchuan Mei and Peizhong Yang and Hongmei Chen and Lizhen Wang}, doi = {10.1016/j.eswa.2025.129775}, journal = {Expert Systems with Applications}, month = {3}, pages = {129775}, shortjournal = {Expert Syst. Appl.}, title = {Mining spatiotemporal dominant co-location patterns}, volume = {298}, year = {2026}, }
• (2026). Cross-scale content adaptive network for three-dimensional multi-object tracking and fish activity quantification. ESWA, 298, 129774. (www)

  @article{ , abstract = {Tracking and quantifying fish activity are vital for evaluating their health status and adaptability to the environment. However, most current research on fish tracking and activity quantification suffers from the limitation of being two-dimensional, losing crucial vertical or horizontal information. To facilitate tracking and quantitative analysis of fish activity in three-dimensional (3D) space, a cross-scale content-adaptive network-based 3D multi-object tracking method for fish is proposed, through which fish movements are quantified accordingly. Firstly, a cross-scale content-adaptive fusion network is proposed to accurately determine the fish positions from top-down and side views, thereby mitigating the issue of scale variation across different perspectives. Secondly, a hierarchical tracking method is implemented to obtain the 3D trajectories of the fish, addressing the challenge of cross-view identity matching. Finally, activity parameters in 3D space, including the activity quantity and trajectory length for individual fish, as well as the dispersion and cohesion for the fish group, are calculated. The proposed method was validated, achieving a Multi-Object Tracking Accuracy (MOTA) of 97.68% and an Identification F1 Score (IDF1) of 97.93%. For activity quantification, the Mean Absolute Error (MAE) was found to be 0.088 (unit weight·(cm/s) 2 ), and the Root Mean Square Error (RMSE) was 0.1064 (unit weight·(cm/s) 2 ). These results affirm the method’s adaption of fish features across scales for 3D tracking and activity analysis. With its efficient performance, our method presents as an instrument for activities such as fish behavior monitoring, selective breeding, and environmental assessment.}, archive = {J_ESWA}, author = {Yiran Liu and Dingshuo Liu and Mingrui Kong and Beibei Li and Qingling Duan}, doi = {10.1016/j.eswa.2025.129774}, journal = {Expert Systems with Applications}, month = {3}, pages = {129774}, shortjournal = {Expert Syst. Appl.}, title = {Cross-scale content adaptive network for three-dimensional multi-object tracking and fish activity quantification}, volume = {298}, year = {2026}, }
• (2026). An adaptive multimodal semantic knowledge enhanced framework for sarcasm detection. ESWA, 298, 129773. (www)

  @article{ , abstract = {Multimodal sarcasm detection (MSD) has become an important research topic for understanding sentiments on social media, while various recent MSD approaches extract high-level semantic knowledge from images to improve performance. However, some key semantic information, such as emotions expressed in images, is still neglected, limiting reliable sentiment understanding. To address this issue, we propose an adaptive multimodal semantic knowledge enhanced framework for sarcasm detection. We first design an adaptive processing pipeline to extract emotion-aware visual semantics as an auxiliary modality to enhance multimodal feature representations. Enabled by two attention mechanisms, bidirectional cross-modal attention and graph attention, interactions between modalities are analysed to improve MSD performance. Extensive experiments are conducted on two public multimodal sarcasm detection datasets, MMSD and MMSD 2.0, comprising approximately 19,000 tweet samples. Our proposed approach achieves consistent improvements in both sarcasm detection accuracy and F1-score compared to strong baseline models such as DIP and KnowleNet. Built upon a ViT-based architecture, the fine-tuned model offers competitive performance with lower computational overhead, highlighting its potential for practical deployment.}, archive = {J_ESWA}, author = {Jing Dong and Yu Sui and Qiang Zhang and Hui Fang and Gerald Schaefer and Rui Liu and Pengfei Yi and Xiaoyong Fang}, doi = {10.1016/j.eswa.2025.129773}, journal = {Expert Systems with Applications}, month = {3}, pages = {129773}, shortjournal = {Expert Syst. Appl.}, title = {An adaptive multimodal semantic knowledge enhanced framework for sarcasm detection}, volume = {298}, year = {2026}, }
• (2026). CrossFPN: Efficient cross-scale feature pyramid network for real-time low-altitude aerial target detection. ESWA, 298, 129772. (www)

  @article{ , abstract = {Low-altitude aerial target detection presents significant challenges in object detection due to the rapid development of Unmanned Aerial Vehicles (UAVs). Despite the recent advances in feature pyramid networks, the accuracy and efficiency of real-time low-altitude aerial target detection remain suboptimal, which is caused by the tiny proportion, the similar patterns, the complex background textures, and the deficient resolution of the target area of interest. This study proposes the innovative Cross-Scale Feature Pyramid Network (CrossFPN) to address the key gap in real-time low-altitude aerial target detection. Benefiting from the core components Deep Cross-Scale Fuser Module (DCFM) and Shallow Cross-Scale Fuser Module (SCFM), the CrossFPN efficiently fuses the semantic and spatial features across levels through a unique propagating route and preserves the high-resolution information that is crucial to small aerial targets. Differing from the conventional feature extraction process, the simultaneous fusion mode and the feature reutilization reduce the computational overhead of the model and ensure real-time performance, uncovering novel insights into cross-scale fusion. Furthermore, the Adaptive Partial Channel Mixer (APCM) is implemented to process the fused features and extract the discriminative features of the low-altitude aerial targets. Publicly available datasets and self-built extended datasets were employed for hybrid training to enhance the robustness and scalability of the model. Extensive experiments on various datasets have demonstrated that CrossFPN performs superior to state-of-the-art feature pyramid networks in detection accuracy and model complexity by a notable margin.}, archive = {J_ESWA}, author = {Liyan Cai and Qingfang Zheng and Yaowei Wang and Xuefei Xiong and Yong Xu}, doi = {10.1016/j.eswa.2025.129772}, journal = {Expert Systems with Applications}, month = {3}, pages = {129772}, shortjournal = {Expert Syst. Appl.}, title = {CrossFPN: Efficient cross-scale feature pyramid network for real-time low-altitude aerial target detection}, volume = {298}, year = {2026}, }
• (2026). FoRKER: Focused reasoner with knowledge editing and self-reflection. ESWA, 298, 129771. (www)

  @article{ , abstract = {Multi-hop question answering (MHQA) is a complex question answering (QA) benchmark that requires agents to integrate information from diverse sources and utilize cross-referencing reasoning to answer intricate questions. Existing MHQA-handling frameworks typically employ a retrieve-read paradigm. However, these efforts rooted in the retrieve-read paradigm are still constrained by: 1) unstable document retrieval performance , 2) weak knowledge refinement capabilities , and 3) the absence of a reflection mechanism for error awareness . To address these limitations, we propose FoRKER ( Fo cused R easoner with K nowledge E diting and Self- R eflection), which is a plug-and-play framework. Specifically, we develop a novel progressive focusing mechanism to pinpoint highly relevant document resources and introduce knowledge editing techniques to further eliminate noise interference within textual information. Additionally, we design a novel prompting method, named Chain-of-Evidence (CoE), which is designed to augment the reasoning capabilities of FoRKER . Notably, the integration of Self-Reflection technology further endows FoRKER with the ability to learn and improve from its mistakes. Extensive experiments on widely-used datasets demonstrate that FoRKER achieves new state-of-the-art results in information retrieval and reading comprehension, while also exhibiting effective generalization. Exhilaratingly, on the MusiqueQA dataset, FoRKER demonstrates a 20 % improvement in Answering scores compared to the advanced competitors.}, archive = {J_ESWA}, author = {Chunbai Zhang and Haoyang Li and Chao Wang and Yang Zhou and Yan Peng}, doi = {10.1016/j.eswa.2025.129771}, journal = {Expert Systems with Applications}, month = {3}, pages = {129771}, shortjournal = {Expert Syst. Appl.}, title = {FoRKER: Focused reasoner with knowledge editing and self-reflection}, volume = {298}, year = {2026}, }
• (2026). Training-free pedestrian trajectory prediction via segmentation-guided path planning. ESWA, 298, 129770. (www)

  @article{ , abstract = {Pedestrian trajectory prediction is a critical research topic for industrial applications and has been significantly advanced by deep learning. Despite decades of progress, current approaches still face two major challenges. First, the scarcity of data limits the generalization capability of deep learning models. Second, the absence of interpretability hinders real-world applications. To address these challenges, recent research has leveraged the knowledge of large language models (LLMs) to alleviate data sparsity and introduced novel knowledge-based methods to enhance interpretability. Nevertheless, transferring LLMs is excessively cumbersome, and the black-box nature of deep learning continues to obstruct interpretability. In our work, we propose a training-free transfer paradigm named Segmentation-Guided Path Planning (SGPP). Rather than directly transferring pretrained LLMs, SGPP introduces a more tailored and efficient transfer strategy by employing a promptable segmentation model. The segmentation model explicitly extracts walkable regions from the scenario, which serve as constraints in the planning space, thereby reformulating trajectory prediction as a more tractable white-box path-planning problem. Within this framework, our method offers a more effective solution to the two prevailing challenges. Compared with the latest training-free methods, our approach achieves superior performance and demonstrates strong generalization across diverse real-world scenarios, highlighting its suitability for industrial applications.}, archive = {J_ESWA}, author = {Dongchen Li and Zhimao Lin and Jinglu Hu}, doi = {10.1016/j.eswa.2025.129770}, journal = {Expert Systems with Applications}, month = {3}, pages = {129770}, shortjournal = {Expert Syst. Appl.}, title = {Training-free pedestrian trajectory prediction via segmentation-guided path planning}, volume = {298}, year = {2026}, }
• (2026). A novel fuzzy clustering approach with transition matrix for explainable evaluation of social media-based digital literacy interventions. ESWA, 298, 129769. (www)

  @article{ , abstract = {Assessing the effectiveness of digital literacy interventions often relies on raw score comparisons or hard classifications, which may obscure nuanced changes in conceptual understanding and provide limited interpretability. Traditional approaches fail to capture the probabilistic and fuzzy nature of learning progression and do not support transparent analysis of how learners transition across conceptual clusters over time. This study proposes an explainable evaluation framework that integrates fuzzy clustering with a fuzzy transition matrix to model the redistribution of aggregated membership values between pretest and posttest conceptual clusters. The framework applies Fuzzy C-Means (FCM) to derive soft cluster memberships and constructs a transition matrix that represents probabilistic learning progression in a linguistically interpretable form. Unlike conventional methods, this approach enables the analysis of gradual transitions across levels of proficiency rather than binary outcomes. The model was applied to real-world educational data from control and experimental classes, the latter of which received a social media-based instructional intervention. Results indicate that the control class exhibited downward or stagnant patterns, particularly among high-performing learners, while the experimental class showed more coherent upward cluster transitions among low- and moderate-level learners. By enabling interpretable modeling of pre–post cluster transition patterns, the proposed framework contributes to the advancement of explainable machine learning in education. It also highlights the potential of social computing platforms to foster scalable, data-driven digital literacy development.}, archive = {J_ESWA}, author = {Rustam and Diana Noor Anggraini and Koredianto Usman and Loveleen Gaur}, doi = {10.1016/j.eswa.2025.129769}, journal = {Expert Systems with Applications}, month = {3}, pages = {129769}, shortjournal = {Expert Syst. Appl.}, title = {A novel fuzzy clustering approach with transition matrix for explainable evaluation of social media-based digital literacy interventions}, volume = {298}, year = {2026}, }
• (2026). Dynamic graph structure correction with nonadjacent correlations for multivariate time series forecasting. ESWA, 298, 129768. (www)

  @article{ , abstract = {Effectively modeling the relations between variables in multivariate time series is of utmost importance for accomplishing accurate predictions. In real-world scenarios, in addition to sequential correlations, the evolution of relations between variables also exhibits nonadjacent correlations at different scales. However, existing methods primarily focus on constructing dynamic graph structures at each time step using temporal features extracted by continuous temporal models, which cannot capture above latent dependencies. In this study, we introduce the Dynamic Graph Structure Correction (DGC) model, leveraging a multi-scale framework with dilated convolution. To take full advantage of nonadjacent correlations in the evolution of relations between variables, we adaptively select history-related graph structures to correct initial graph structure constructed by Gate Recurrent Units. In addition, we design a time-decay-based attention mechanism to address the influence of time intervals between history-related and current time steps. Finally, the evolved graph structures are fed into graph neural networks to handle the multi-scale and complex structural relations. Our proposed model achieves superior performance compared to state-of-the-art methods in multivariate time series forecasting, as evidenced by the evaluation results on four widely used benchmark datasets.}, archive = {J_ESWA}, author = {Dandan He and Yueyang Wang and Chaoli Lou and Gang Tan and Qingyu Xiong and Guodong Sa}, doi = {10.1016/j.eswa.2025.129768}, journal = {Expert Systems with Applications}, month = {3}, pages = {129768}, shortjournal = {Expert Syst. Appl.}, title = {Dynamic graph structure correction with nonadjacent correlations for multivariate time series forecasting}, volume = {298}, year = {2026}, }
• (2026). Optimized homomorphic linear computation in privacy-preserving CNN inference. ESWA, 298, 129767. (www)

  @article{ , abstract = {Machine Learning as a Service (MLaaS) provides robust solutions for deploying deep learning inference in cloud environments. However, it also raises serious privacy concerns regarding user data and proprietary model parameters. Numerous hybrid cryptographic protocols that integrate homomorphic encryption (HE) and garbled circuits (GC) have been proposed to enable secure inference with low latency. In these protocols, the homomorphic evaluation of linear operations remains the primary performance bottleneck and warrants further optimization. In this work, we propose novel optimizations for HE-based linear computations within the hybrid cryptographic framework for secure neural network inference. Specifically, we devise two efficient strategies for homomorphic matrix-vector multiplication and convolution. For matrix-vector multiplication, we introduce a grouped diagonal extraction technique that encodes the weight matrix more compactly and enables configurable ciphertext rotation reuse, while for homomorphic convolution, we present a group-wise combine-and-merge evaluation method. Both methods significantly reduce the number of required ciphertext rotations. Our approach achieves up to a 3.9 × speedup in matrix-vector multiplication and a 2.9 × improvement in convolution over state-of-the-art (SOTA) solutions. The HE-GC hybrid secure convolutional neural networks (CNN) inference framework incorporating these enhancements yields speedups of 2.5 × on widely used ResNets deep learning architectures.}, archive = {J_ESWA}, author = {Chenglong Li and Xirong Ma and Xiuhao Wang and Fanyu Kong and Yunting Tao and Chunpeng Ge}, doi = {10.1016/j.eswa.2025.129767}, journal = {Expert Systems with Applications}, month = {3}, pages = {129767}, shortjournal = {Expert Syst. Appl.}, title = {Optimized homomorphic linear computation in privacy-preserving CNN inference}, volume = {298}, year = {2026}, }
• (2026). SS-RDHEI for embedding capacity enhancement by PDPM and auxiliary data free coding. ESWA, 298, 129766. (www)

  @article{ , abstract = {Reversible data hiding in encrypted images (RDHEI) is a promising technique for multimedia cloud computing that enables the embedding of secret data into encrypted images while preserving confidentiality. However, the existing RDHEI algorithms fail to meet the high-security requirements of distributed storage systems in the cloud. Although, secret sharing based RDHEI (SS-RDHEI) may solve this problem, the current methods have weakness such as insufficient embedding capacity and unsatisfactory balance between image security and redundancy. To enhance the algorithm’s ability to carry information, this paper proposes a SS-RDHEI for embedding capacity enhancement by PDPM and auxiliary data free coding. Firstly, pixel-difference preservation based modulation (PDPM) ensures secure encryption by modifying all pixels except for a reference block, minimizing damage; moreover, an improved block-level pixel predictor enhances carrier redundancy. Secondly, auxiliary data free coding (ADFC) marks prediction errors directly in the binary sequence of the original pixel without auxiliary information while maintaining accuracy, and reduces the impact of different textures on embedding performance byselecting optimal parameters for each share image. Finally, by combining PDPM with secret sharing, it achieves independent embedding for multiple data hiders while ensuring fair information embedding. Experimental results demonstrate that the proposed algorithm outperforms existing state-of-the-art schemes in terms of information-carrying capability.}, archive = {J_ESWA}, author = {Zhihua Gan and Zongwei Tang and Yalin Song and Gongyao Cao and Xiuli Chai and Yushu Zhang}, doi = {10.1016/j.eswa.2025.129766}, journal = {Expert Systems with Applications}, month = {3}, pages = {129766}, shortjournal = {Expert Syst. Appl.}, title = {SS-RDHEI for embedding capacity enhancement by PDPM and auxiliary data free coding}, volume = {298}, year = {2026}, }
• (2026). Improved convolutional neural networks for the bullwhip effect in supply chains. ESWA, 298, 129764. (www)

  @article{ , abstract = {The Bullwhip Effect (BWE) introduces significant challenges to production systems by amplifying demand and order oscillations. One of the most effective methods for predicting and modeling complex systems is Convolutional Neural Networks (CNNs). However, certain phenomena, such as the BWE in supply chains (SC), are difficult to predict and identify directly. The primary challenge for Machine Learning (ML) algorithms in this context lies in the training phase: the raw demand and order data are fed into the network, yet the desired training outcome is the oscillatory behavior of these data from the perspective of the BWE. Consequently, conventional max pooling, average pooling operators, kernels, and weighted linear combinations of data are insufficient for capturing this type of learning. To address this issue, in this paper, a novel structure containing new pooling operators and kernels of CNNs is proposed to tailor the unique characteristics of the BWE. Specifically: a ) Considering the temporal propagation nature of the BWE, new filters and pooling operators were designed to enable CNNs to predict the BWE accurately. b ) A tensor structure was also proposed for the time signal of demand as inputs of the CNNs to facilitate the analysis of all factors influencing the occurrence of the BWE. c ) To capture the magnitude of the BWE among features, a novel combination of filters and pooling operators was proposed, enabling the CNNs to account for hidden but yet significant feature effects during training. The benefits of the proposed approach lie in its versatility, and it can be applied to train CNNs to model structured fluctuations like the BWE in various dynamic systems.}, archive = {J_ESWA}, author = {Sajjad Aslani Khiavi and Farzad Hashemzadeh}, doi = {10.1016/j.eswa.2025.129764}, journal = {Expert Systems with Applications}, month = {3}, pages = {129764}, shortjournal = {Expert Syst. Appl.}, title = {Improved convolutional neural networks for the bullwhip effect in supply chains}, volume = {298}, year = {2026}, }
• (2026). A hybrid risk assessment method combining CatBoost and FAHP-grid search optimized risk matrix for container ship accident. ESWA, 298, 129763. (www)

  @article{ , abstract = {As a dominant mode of maritime transportation with unique risk characteristics, container shipping requires accurate and applicable risk assessment. However, conventional risk matrices oversimplify complex interactions, while pure data-driven models lack operational utility. To address this, a hybrid method for container ship risk assessment is proposed. This method integrates CatBoost-based predictive method, FAHP-grid search optimized risk matrix, and GIS-supported risk mapping. A comprehensive study of maritime casualties and piracy accidents is conducted, utilizing historical incident data sets collected from the Global Integrated Shipping Information System (GISIS). The global maritime accident risk of container ships is then evaluated and mapped. The sensitivity analysis confirms the robustness of the method under varying linguistic distance parameters, while expert weights have a moderate impact on the assessment results. Finally, the effectiveness of the proposed method is validated through comparative analyses on predictive performance, risk discrimination capability, and risk assessment accuracy. CatBoost algorithm outperforms XGBoost, LightGBM, and Random Forest algorithms in predictive metrics. The designed risk matrix shows strong discriminatory ability for container ship risk levels. In historical accident data validation, the proposed method also achieves higher accuracy than combinations involving XGBoost, LightGBM, or Random Forest with the designed risk matrix.}, archive = {J_ESWA}, author = {Yuqing Xiao and Shilian Han and Xinwang Liu}, doi = {10.1016/j.eswa.2025.129763}, journal = {Expert Systems with Applications}, month = {3}, pages = {129763}, shortjournal = {Expert Syst. Appl.}, title = {A hybrid risk assessment method combining CatBoost and FAHP-grid search optimized risk matrix for container ship accident}, volume = {298}, year = {2026}, }
• (2026). Quadrotor navigation considering attitude: A deep reinforcement learning method using tangent path rewards. ESWA, 298, 129762. (www)

  @article{ , abstract = {Autonomous navigation of quadrotors is a fundamental prerequisite for numerous applications. This work proposes a novel deep reinforcement learning (DRL) framework that explicitly addresses quadrotor attitude dynamics during autonomous navigation, a critical yet underexplored challenge in existing learning-based UAV navigation studies. In the proposed method, high-level velocity commands will be generated by a deep neural network policy and translated by a low-level control algorithm to achieve precise control of both positions and rotations of quadrotors. A specialized network structure is designed to effectively extract environmental obstacle features and quadrotor sequence features to improve navigation performance. In addition, a novel tangent path reward (TPR) calculation method is developed to adequately utilize the known contours and positions of obstacles during the training phase. Experimental results demonstrate that the proposed method enables quadrotors to autonomously navigate complex virtual obstacle environments with superior efficiency compared with other algorithms. Furthermore, the feasibility and adaptability of the proposed method are validated through simulations by varying obstacle density and map size, as well as replicating real-world obstacle distributions.}, archive = {J_ESWA}, author = {Qizhang Luo and Yuqi Li and Jiaheng Zeng and Guohua Wu and Yalin Wang}, doi = {10.1016/j.eswa.2025.129762}, journal = {Expert Systems with Applications}, month = {3}, pages = {129762}, shortjournal = {Expert Syst. Appl.}, title = {Quadrotor navigation considering attitude: A deep reinforcement learning method using tangent path rewards}, volume = {298}, year = {2026}, }
• (2026). A dynamic tri-stage framework with neural network-assisted search for constrained multi-objective optimization. ESWA, 298, 129761. (www)

  @article{ , abstract = {Constrained multi-objective optimization problems involve the optimization of multiple objective functions and the satisfaction of different constraints, which poses a challenge for algorithms to achieve a good balance between convergence and diversity. However, indiscriminately enhancing diversity can hinder convergence, while solely focusing on convergence may impair the exploration of the objective space, especially when the current stage is not well-defined. To address this issue, we propose a three-stage multi-task framework for constrained multi-objective optimization with dynamically switchable stages. This framework introduces two auxiliary tasks: one that operates during the exploration and transition stages to accelerate convergence towards the boundary of the infeasible regions and assist the population in crossing it, and another that operates in the final convergence stage to guide the population towards the constrained Pareto front. Moreover, a stage detection method is proposed, which evaluates the current stage to determine the appropriate evolutionary direction for the population, thus enabling dynamic stage transitions. In addition, a neural network-assisted search operator is designed for the auxiliary task during the transition stage, which learns the optimal offspring generation process. This operator enhances the ability of the auxiliary population to cross the infeasible regions. Finally, the performance of the proposed algorithm is superior and competitive on three test suites and six real-world engineering problems compared to seven state-of-the-art algorithms.}, archive = {J_ESWA}, author = {Qianlong Dang and Xinkang Hong and Xianpeng Sun}, doi = {10.1016/j.eswa.2025.129761}, journal = {Expert Systems with Applications}, month = {3}, pages = {129761}, shortjournal = {Expert Syst. Appl.}, title = {A dynamic tri-stage framework with neural network-assisted search for constrained multi-objective optimization}, volume = {298}, year = {2026}, }
• (2026). A physics-informed neural network surrogate model and many-objective optimization algorithm for coupled multi-energy systems in smart grids. ESWA, 298, 129760. (www)

  @article{ , abstract = {The scope of smart grids is progressively extending toward Integrated Energy Systems (IES) that couple electricity with gas, heating, and cooling. Due to the unsteady-state physical characteristics inherent in the transmission of gas, heat, and cooling resources, IES scheduling must not only balance multiple typical objectives but also account for the dynamic coupling of heterogeneous physical domains. To address these challenges, this paper formulates a Many-objective Optimization Model for Coupled Multi-Energy Flows (MaOCMFM) with partial differential equations (PDEs) in IES, which captures the dynamic physical behaviors of electricity, gas, heat, and cooling subsystems. Building upon this model, we propose a Probabilistic Contributing Many-objective Evolutionary Algorithm enhanced by a Physics-Informed Neural Network surrogate model (PC-MaOEA-PINN). Cubic B-spline functions are employed to achieve a continuous representation of the decision variables, while multi-physics constraints are embedded into the loss function of the surrogate model. This design enables efficient approximation of the objective function with a limited number of samples and facilitates focused exploration in critical evolutionary regions, thereby accelerating population convergence. The effectiveness of the proposed model and algorithm is validated on 9 typical scheduling days across four simulated IES scenarios.}, archive = {J_ESWA}, author = {Jingbo Zhang and Xingjuan Cai and Zhihua Cui and Jinjun Chen}, doi = {10.1016/j.eswa.2025.129760}, journal = {Expert Systems with Applications}, month = {3}, pages = {129760}, shortjournal = {Expert Syst. Appl.}, title = {A physics-informed neural network surrogate model and many-objective optimization algorithm for coupled multi-energy systems in smart grids}, volume = {298}, year = {2026}, }
• (2026). An emergency scheduling method based on AutoML for space maneuver objective tracking. ESWA, 298, 129759. (www)

  @article{ , abstract = {The rapid development of large-scale constellations has led to a dramatic increase in the number of Resident Space Objectives (RSOs), significantly intensifying the complexity of Space Situational Awareness (SSA). Furthermore, the maneuver behaviors of non-cooperative RSOs pose potential threats to space safety, making the real-time monitoring of their post-maneuver orbital becomes more critical. In particular, the maneuvering characteristics of large-scale constellation satellites impose more stringent demands on the timeliness and adaptability of existing scheduling algorithms for observation resources. To address the emergency scheduling demands of heterogeneous ground-based observation resources, this paper proposes an Emergency Task Three-phase Scheduling Framework (ETTSF) based on Automated Machine Learning (AutoML) and auction algorithm. The framework collaboratively optimizes resource allocation through three phases: resource matching, task scheduling, and rescheduling. First, AutoML combined with an auction algorithm predicts and assigns emergency tasks to the most appropriate resources, reducing solution space complexity, simultaneously, the auction algorithm’s corrected results are fed back to AutoML for model fine-tuning. Second, a heuristic algorithm with dynamic neighborhood structures efficiently inserts emergency tasks into the routine observation plan. Finally, affected routine tasks are rescheduled to minimize the operational impact. Simulation results demonstrate that compared to baselines-Real-Time Dynamic Scheduling (RTDS) and Improved Adaptive Large Neighborhood Search (IALNS)-ETTSF achieves a 2.82% improvement in Completion Rate of Emergency RSOs (CRER) and a 27.83% reduction in Impact Rate (IR) on routine tasks. Ablation experiments further validate the effectiveness of the resource matching and rescheduling phases.}, archive = {J_ESWA}, author = {Xi Long and Jinrun Chen and Leping Yang and Huan Huang}, doi = {10.1016/j.eswa.2025.129759}, journal = {Expert Systems with Applications}, month = {3}, pages = {129759}, shortjournal = {Expert Syst. Appl.}, title = {An emergency scheduling method based on AutoML for space maneuver objective tracking}, volume = {298}, year = {2026}, }
• (2026). Exploiting point-language models with dual-prompts for 3D anomaly detection. ESWA, 298, 129758. (www)

  @article{ , abstract = {Anomaly detection (AD) in 3D point clouds is crucial in a wide range of industrial applications, especially in various forms of precision manufacturing. Considering the industrial demand for reliable 3D AD, several methods have been developed. However, most of these approaches typically require training separate models for each category, which is memory-intensive and lacks flexibility. In this paper, we propose a novel P oint- L anguage model with dual-prompts for 3D AN omaly d E tection (PLANE). The approach leverages multi-modal prompts to extend the strong generalization capabilities of pre-trained Point-Language Models (PLMs) to the domain of 3D point cloud AD, achieving impressive detection performance across multiple categories using a single model. Specifically, we propose a dual-prompt learning method, incorporating both text and point cloud prompts. The method utilizes a dynamic prompt creator module (DPCM) to produce instance-specific dynamic prompts, which are then integrated with class-specific static prompts for each modality, effectively driving the PLMs. Additionally, based on the characteristics of point cloud data, we propose a pseudo 3D anomaly generation method (Ano3D) to improve the model’s detection capabilities in the unsupervised setting. Experimental results demonstrate that the proposed method, which is under the multi-class-one-model paradigm, achieves a +8.7 %/+7.0 % gain on anomaly detection and localization performance as compared to the state-of-the-art one-class-one-model methods for the Anomaly-ShapeNet dataset, and obtains +4.3 %/+0.3 % gain for the Real3D-AD dataset. Code will be available upon publication.}, archive = {J_ESWA}, author = {Jiaxiang Wang and Haote Xu and Xiaolu Chen and Haodi Xu and Yue Huang and Xinghao Ding and Xiaotong Tu}, doi = {10.1016/j.eswa.2025.129758}, journal = {Expert Systems with Applications}, month = {3}, pages = {129758}, shortjournal = {Expert Syst. Appl.}, title = {Exploiting point-language models with dual-prompts for 3D anomaly detection}, volume = {298}, year = {2026}, }
• (2026). Path-aware routing system for multimodal vigilance estimation: A structured fusion perspective. ESWA, 298, 129757. (www)

  @article{ , abstract = {Driver vigilance is a critical cognitive factor in ensuring the safety of intelligent driving systems. With advances in physiological and behavioral sensing technologies, multimodal data have become an essential source for modeling drivers’ cognitive states. However, vigilance, as an implicit cognitive state, is difficult to model accurately using a single modality. Efficient fusion of multiple modalities and structured information interaction remains a core challenge in this task. Existing approaches often rely on strategies such as feature concatenation and attention mechanisms, which lack explicit structural constraints. As a result, heterogeneous modality features are fused in an uncontrolled and entangled manner, making the interaction process opaque and difficult to interpret. To address these issues, we propose the Path-Aware Routing System (PARS), which formulates multimodal fusion as a feature routing task. In PARS, intra-modal enhancement and cross-modal interaction are abstracted into independent semantic channels, and a confidence-aware mechanism is introduced to enable dynamically weighted fusion. PARS explicitly constructs a path space, allowing features to be selectively routed and integrated along semantical pathways, thereby enhancing the model’s discriminative power and robustness. We conduct extensive experiments on a public dataset and a self-constructed driving simulation dataset. The results demonstrate that PARS significantly outperforms existing multimodal fusion methods in the task of driver vigilance estimation, achieving superior performance in terms of accuracy, interpretability, and generalization. These findings highlight the broad potential of PARS in intelligent driving applications. Our code and models are available at: https://github.com/SunYu-Gavin/PARS .}, archive = {J_ESWA}, author = {Yu Sun and Shiwu Li and Yiming Bie and Linhong Wang and Tongtong Jin and Mengzhu Guo and Zhifa Yang}, doi = {10.1016/j.eswa.2025.129757}, journal = {Expert Systems with Applications}, month = {3}, pages = {129757}, shortjournal = {Expert Syst. Appl.}, title = {Path-aware routing system for multimodal vigilance estimation: A structured fusion perspective}, volume = {298}, year = {2026}, }
• (2026). Homophone-aware offensive language detection via semantic-phonetic collaboration. ESWA, 298, 129756. (www)

  @article{ , abstract = {The increasing use of implicit and obfuscated expressions poses significant challenges to offensive language detection in Chinese online platforms. In particular, users often exploit homophone substitutions to bypass keyword-based moderation, making traditional detection systems inadequate. This study addresses the problem of detecting offensive content masked through homophonic substitutions, which retain aggressive intent while altering character representations. Existing methods fall into two main categories: (1) semantic-only models, which struggle with phonetic manipulations due to their reliance on text features alone, and (2) auxiliary-enhanced models, which incorporate phonetic or syntactic signals but lack deep integration between modalities. To overcome these limitations, we propose a lightweight dual-branch model that separately encodes textual semantics and pinyin phonetics under a multi-view learning framework. A Dual-Branch Interactive Training strategy is introduced to enable dynamic cross-modal alignment via contrastive objectives, allowing each modality to mutually refine the other and enhance robustness to adversarial inputs. We conduct experiments on two benchmark datasets, COLD and SWSR, both of which are augmented with varying levels of homophone noise to simulate real-world evasion strategies. The proposed model outperforms all baseline models, achieving an average F1-score improvement of 6.3 % under high-noise conditions, while reducing inference latency and memory usage by more than 60 %, demonstrating both effectiveness and efficiency for real-time deployment. We will release the source code for further use by the community https://github.com/hjhhlc/DBIT .}, archive = {J_ESWA}, author = {Jiahao Hu and Shanliang Pan}, doi = {10.1016/j.eswa.2025.129756}, journal = {Expert Systems with Applications}, month = {3}, pages = {129756}, shortjournal = {Expert Syst. Appl.}, title = {Homophone-aware offensive language detection via semantic-phonetic collaboration}, volume = {298}, year = {2026}, }
• (2026). Enhancing traffic signal control through model-based reinforcement learning and policy reuse. ESWA, 298, 129755. (www)

  @article{ , abstract = {Multi-agent reinforcement learning (MARL) has shown significant potential in traffic signal control (TSC). However, current MARL-based methods often suffer from insufficient generalization due to the fixed traffic patterns and conditions of the road network used during training. This limitation results in poor adaptability to new traffic scenarios, leading to high retraining costs and complex deployment. To address this challenge, we propose two algorithms: PLight and PRLight. PLight employs a model-based reinforcement learning approach, pretraining control policies, and environment models using predefined source-domain traffic scenarios. The environmental model predicts state transitions, facilitating the comparison of environmental characteristics. PRLight further enhances adaptability by adaptively selecting pre-trained PLight agents based on the similarity between the source and target domains to accelerate the learning process in the target domain. We evaluated the algorithms through two transfer settings: (1) adaptability to different traffic scenarios within the same road network, and (2) generalization across different road networks. The results show that PRLight significantly reduces the adaptation time compared to learning from scratch in new TSC scenarios, achieving optimal performance using similarities between available and target scenarios.}, archive = {J_ESWA}, author = {Yihong Li and Chengwei Zhang and Furui Zhan and Wanting Liu and Kailing Zhou and Longji Zheng}, doi = {10.1016/j.eswa.2025.129755}, journal = {Expert Systems with Applications}, month = {3}, pages = {129755}, shortjournal = {Expert Syst. Appl.}, title = {Enhancing traffic signal control through model-based reinforcement learning and policy reuse}, volume = {298}, year = {2026}, }
• (2026). Federated transfer learning for anomaly detection in HPC systems: First real-world validation on a tier-0 supercomputer. ESWA, 298, 129754. (www)

  @article{ , abstract = {High-Performance Computing (HPC) systems increasingly require intelligent, scalable anomaly detection to ensure operational reliability. However, conventional centralized approaches often struggle with data privacy constraints, poor generalization across heterogeneous nodes, and limited scalability. This study presents the first real-world application of federated transfer learning (FTL) for anomaly detection in a production-grade Tier-0 supercomputer. By combining federated learning with transfer learning, the proposed framework enables decentralized model training and personalized adaptation to unseen nodes, without accessing raw data. We validate the approach using two large-scale telemetry datasets collected from 100 nodes of the Marconi100 supercomputer, evaluating its effectiveness across supervised, semi-supervised, and unsupervised learning paradigms. Results show that FTL consistently improves anomaly detection performance on nodes that did not participate in federated training, with F1-score gains reaching up to 0.50. These improvements demonstrate the framework’s ability to generalize across non-identically distributed data and maintain detection accuracy under real-world conditions. This work establishes FTL as a scalable, privacy-preserving solution for fault detection in HPC environments. Its practical deployment on production hardware confirms its readiness for real-time monitoring applications in large-scale, heterogeneous computing systems.}, archive = {J_ESWA}, author = {Emmen Farooq and Michela Milano and Andrea Borghesi}, doi = {10.1016/j.eswa.2025.129754}, journal = {Expert Systems with Applications}, month = {3}, pages = {129754}, shortjournal = {Expert Syst. Appl.}, title = {Federated transfer learning for anomaly detection in HPC systems: First real-world validation on a tier-0 supercomputer}, volume = {298}, year = {2026}, }
• (2026). Security script arrangement based on enhanced BERT for cooperative defense in networked control systems. ESWA, 298, 129753. (www)

  @article{ , abstract = {Due to the mutual collaboration and in-depth integration among multiple defense technologies through information sharing, the cooperative defense in networked control systems has emerged as a feasible solution to counter increasingly diversified cyber threats under the unique security characteristics and requirements of industrial environments. However, one of the chief challenges is how to automatically and intelligently develop effective cooperative working strategies when an attack occurs. Leveraging the advantages of large-scale AI (Artificial Intelligence) models, this paper defines a new concept named “security script”, and proposes a security script arrangement approach based on enhanced BERT to achieve fine-grained cooperative defense in networked control systems. Furthermore, this approach introduces intrusion detection and industrial firewall as two practical examples, and can automatically arrange effective security scripts to enable the dynamic interaction of two defense technologies. Additionally, to improve efficiency, the encoder structure adjusting and AdamW optimizing are further presented to enhance the traditional BERT. Experimental results clearly demonstrate that: for one thing, these two optimization ways can make greater achievements in reducing unnecessary time consumption and enhancing accuracy of security script arrangement; for another, compared with other typical BERT and large-scale AI models, the proposed approach can exhibit more favorable performance advantages in achieving cooperative defense based on security script arrangement. In particular, through its successful application and verification in one real-world manufacturing control system, our approach may bring a potential opportunity or direction for further research and improvement of AI-based cooperative defense.}, archive = {J_ESWA}, author = {Ming Wan and Xueqing Liu and Shengbao An and Aiping Tan and Xi Jin and Chuan Sheng}, doi = {10.1016/j.eswa.2025.129753}, journal = {Expert Systems with Applications}, month = {3}, pages = {129753}, shortjournal = {Expert Syst. Appl.}, title = {Security script arrangement based on enhanced BERT for cooperative defense in networked control systems}, volume = {298}, year = {2026}, }
• (2026). Anatomically-robust and feature-unbiased domain generalization for medical segmentation. ESWA, 298, 129752. (www)

  @article{ , abstract = {Single-source domain generalization (SSDG) in medical image segmentation is severely hindered by limited single-source data and inter-domain biases. Existing methods often focus on appearance bias, neglecting crucial organ shape bias arising from differing acquisition protocols or patient populations and feature bias arising from learned confounding context by network. Therefore, we develop an anatomically-robust and feature-unbiased (ARFU) framework. Specifically, to generate morphologically sound and stylistically diverse training data, ARFU synergistically employs shape regularization-guided augmentation (SRG) and anatomical prior-guided augmentation (APG). The SRG leverages intrinsic structural information from low-frequency as a regularization to guide global appearance transformations while preventing anatomical distortion. APG learns a universal organ shape distribution from the variety of training masks, while simultaneously using organ-specific appearance augmentation to enforce an appearance-agnostic anatomical discrimination. To ensure that the network focuses on discriminative and domain-invariant representations of anatomical features, the feature unbiased learning (FUL) module is designed. FUL perturbs feature distributions by treating intra-batch variance as pixel-wise uncertainty to reduce the model’s reliance on domain-specific cues, and then dynamically filters in the frequency domain to adaptively suppress domain-specific noise and irrelevant background patterns. Extensive experiments on abdominal cross-modality and cardiac cross-sequence segmentation tasks demonstrate that our ARFU framework achieves superior generalization performance compared to state-of-the-art methods.}, archive = {J_ESWA}, author = {Bijuan Ren and Yanfeng Li and Jia Sun and Houjin Chen and Luyifu Chen}, doi = {10.1016/j.eswa.2025.129752}, journal = {Expert Systems with Applications}, month = {3}, pages = {129752}, shortjournal = {Expert Syst. Appl.}, title = {Anatomically-robust and feature-unbiased domain generalization for medical segmentation}, volume = {298}, year = {2026}, }
• (2026). ScSCDT: Self-contrastive neural network with deep topology mining for scRNA-seq data clustering. ESWA, 298, 129751. (www)

  @article{ , abstract = {Advancements in single-cell sequencing technologies have enabled researchers to better identify cells based on gene-level information. Cell clustering is a key task in single-cell analysis and plays an important role in distinguishing cell types. However, due to the high dimensionality and sparsity of scRNA-seq data, single-cell clustering remains a major challenge. Although many methods based on deep learning and machine learning have been developed for single-cell clustering, they often fail to capture the deep topological structure between cells, which limits clustering precision. In addition, most existing clustering approaches cannot effectively construct suitable sample pairs to optimize clustering models. To address these issues, we propose a topology-aware deep contrastive clustering model for single-cell data, named scSCDT. First, scSCDT employs a ZINB-based autoencoder to simultaneously learn cell embeddings and topological information, effectively handling the challenges posed by the high-dimensional and sparse nature of the data. Then, we introduce a dual clustering-guided loss to supervise the clustering task, combining a probabilistic soft assignment strategy and a hard pseudo-labeling strategy for optimization. Finally, based on the topological structure in the low-dimensional embedding space, we construct negative pairs within a single view and design a self-contrastive learning method to further improve clustering performance. We conduct extensive experiments on ten real scRNA-seq datasets and evaluate performance using four clustering metrics. The results indicate that scSCDT achieves strong clustering performance across multiple datasets, thereby facilitating more accurate cell type identification in single-cell transcriptomic analysis.}, archive = {J_ESWA}, author = {Zhongyang Zhou and Bin Tang and Feiyu Chen and Wei Wang and Shangshang Zhao and Nanjun Yu}, doi = {10.1016/j.eswa.2025.129751}, journal = {Expert Systems with Applications}, month = {3}, pages = {129751}, shortjournal = {Expert Syst. Appl.}, title = {ScSCDT: Self-contrastive neural network with deep topology mining for scRNA-seq data clustering}, volume = {298}, year = {2026}, }
• (2026). Visual mamba-CNN for scribble-based segmentation in weakly supervised learning for photoacoustic tomography. ESWA, 298, 129749. (www)

  @article{ , abstract = {Photoacoustic (PA) imaging is a powerful non-invasive medical imaging technique that combines the high contrast of optical imaging with the deep tissue penetration of ultrasound, offering both structural and functional insights into tissues and organs. Organ-level analysis of photoacoustic tomography (PAT) images enables quantification of specific morphological and functional parameters, making accurate organ segmentation a critical step in PA image-based analysis. However, the limited availability of large-scale annotated datasets remains a major challenge. To address this, we employ cross-modality data augmentation by generating synthetic PA images from MRI scans. To further reduce manual annotation efforts, we propose a weakly supervised learning (WSL) framework that leverages scribble annotations. Since many deep learning models struggle to capture global context from sparse labels, we introduce a novel architecture that combines traditional convolutional neural networks (CNNs) with Visual Mamba, integrating both local and global feature extraction capabilities. This hybrid design improves segmentation performance in weakly supervised settings. We validate our method on a simulated PA abdominal dataset and real in vivo mouse abdominal PAT data, demonstrating notable improvements in segmentation accuracy and robustness.}, archive = {J_ESWA}, author = {Meng Zhou and Ziyin Ren and Qinlin Tan and Xin Du and Hengrong Lan and Fei Gao and Raymond Kai-Yu Tong}, doi = {10.1016/j.eswa.2025.129749}, journal = {Expert Systems with Applications}, month = {3}, pages = {129749}, shortjournal = {Expert Syst. Appl.}, title = {Visual mamba-CNN for scribble-based segmentation in weakly supervised learning for photoacoustic tomography}, volume = {298}, year = {2026}, }
• (2026). 3D modeling from a single sketch with multifaceted semantic understanding. ESWA, 298, 129748. (www)

  @article{ , abstract = {This paper studies the problem of 3D shape generation from a single sketch. Prior works rely on directly extracted visual features of sketches as guidance for the generation process. However, the sparse visual cues and abstract nature of sketches, which are inherited in the guiding features, lead to semantic ambiguity and geometry incompleteness in the generated shapes, compromising accuracy. To address this, we propose MSU-3D, a diffusion-based framework for sketch-to-3D generation, leveraging Multifaceted Semantic Understanding to explicitly analyze the construction information of sketches from multiple facets before providing fine-grained guidance over 3D shape generation. Specifically, we decompose sketches through three interpretative facets (semantics, depth, and normal), introducing reasoning of three representations to capture 3D features from distinct perspectives: local components, basic 3D geometry, and 3D surface details. One step further, we propose a multifaceted perception module. It aggregates multifaceted feature representations and leverages local component features as a two-pronged guiding representation to jointly guide the perception of basic shapes and surface details. To ensure fine-grained control, the hierarchical perception strategy adaptively injects varying granularity of perception features at different stages of the 3D generation. Extensive experiments and comparisons with state-of-the-art methods on various complex posture datasets validate the effectiveness of our framework in mitigating semantic ambiguity and geometry incompleteness in 3D generation.}, archive = {J_ESWA}, author = {Yuxiao Zhang and Jin Wang and Yang Zhou and Senyun Jia and Zhi Zheng and Dongliang Zhang and Guodong Lu}, doi = {10.1016/j.eswa.2025.129748}, journal = {Expert Systems with Applications}, month = {3}, pages = {129748}, shortjournal = {Expert Syst. Appl.}, title = {3D modeling from a single sketch with multifaceted semantic understanding}, volume = {298}, year = {2026}, }
• (2026). Continuous learning approach to synergize shimmering image enhancement and de-fogging with small sample. ESWA, 298, 129747. (www)

  @article{ , abstract = {Shimmering Image Enhancement and Defogging (SIED) are two important aspects of image recovery. However, most methods are often fail to consider image context information, overexposure of image and the intrinsic correlation between SIED without enough sample. Furthermore, current methods may lead to the amplification of external color interference during image recovery, and can’t fine-tune the model with new samples. Firstly, we propose an Adjustable Multiscale Attention Codec Network (AMACNet) architecture. AMACNet includes variable restorative coder decoder and group channel attention to fuse multi-level contextual and channel information of images, respectively. These components are designed as plug-and-play modules. Secondly, a Continuous Learning with Small Sample (CLS) training method is presented. It utilizes few samples based on a front-and-back stage receding structure. The method synergizes the tasks of SIED, allowing the network to perform both tasks without incresing the number of parameters. It also enables the network to adapt to new tasks with a small number of unpaired samples. Finally, a color adjustment module for image post-processing is designed. The post-processing method is used to balance the impact of color illumination on the inherent color of materials. Extensive experiments demonstrate that training AMACNet by CLS allow the final image recovery results to exceed the GT in terms of ENIQA, FES, PIQE, and other unparameterized metrics.}, archive = {J_ESWA}, author = {Fenglin Yao and Zhengxiang Liu}, doi = {10.1016/j.eswa.2025.129747}, journal = {Expert Systems with Applications}, month = {3}, pages = {129747}, shortjournal = {Expert Syst. Appl.}, title = {Continuous learning approach to synergize shimmering image enhancement and de-fogging with small sample}, volume = {298}, year = {2026}, }
• (2026). Learning domain-invariant representation for generalizable iris segmentation. ESWA, 298, 129746. (www)

  @article{ , abstract = {Cross-domain iris segmentation (CDIS) seeks to transfer knowledge from a labeled source dataset to an unlabeled target dataset. Existing CNN-based iris segmentation methods commonly assume that training and application stages share the same data distribution and modality setting, thus their performance may decline substantially on open-domain iris datasets unseen before. Furthermore, the process of annotating pixel-wise labels is labor-intensive and time-consuming, resulting in limited applicability of these methods in realistic scenarios. Therefore, we propose a generic domain adaptation iris segmentation framework ( DAIrisSeg ), which can be flexibly incorporated into existing methods. First, a domain-sensitive feature whitening strategy is proposed to effectively mitigate the domain-specific styles while preserving the domain-invariant content, thereby improving the model’s generalizability to unknown domain distribution. We then utilize the prototype estimation and the context-similarity learning adapter to produce reliable segmentation labels. In addition, DAIrisSeg incorporates prior constraints of the iris to further refine the segmentation results. Extensive experiments on three iris datasets demonstrate that the proposed method has shown consistent improvements over state-of-the-art (SOTA) methods.}, archive = {J_ESWA}, author = {Dawei Lin and Meng Yuan and Ying Chen and Xiaodong Zhu and Yuanning Liu}, doi = {10.1016/j.eswa.2025.129746}, journal = {Expert Systems with Applications}, month = {3}, pages = {129746}, shortjournal = {Expert Syst. Appl.}, title = {Learning domain-invariant representation for generalizable iris segmentation}, volume = {298}, year = {2026}, }
• (2026). Cycle-CFM: An unsupervised framework for robust multimodal anomaly detection in industrial settings. ESWA, 298, 129745. (www)

  @article{ , abstract = {Industrial multimodal anomaly detection is confronted with three pivotal challenges: cross-modal feature drift, noise sensitivity, and modality imbalance. To address these issues, we propose Cycle-Consistent Cross-Modal Feature Mapping (Cycle-CFM), an unsupervised framework that integrates cycle-consistent cross-modal mapping with channel-attention-guided adaptive loss weighting. Cycle-CFM establishes bidirectional feature alignment between RGB and 3D modalities via reversible cycle mappings, yielding consistent representations robust to vibration and depth noise. To further mitigate dynamic interferences such as illumination variations, we introduce a joint optimization strategy that combines cross-consistency and cycle-consistency losses. Experimental results on our self-constructed SteelDefect-3D-AD dataset demonstrate that Cycle-CFM achieves an AUPRO@1 % of 0.371, outperforming state-of-the-art methods by 17–45 %. It also attains a pixel-level AUROC (P-AUROC) of 0.991 and an image-level AUROC (I-AUROC) of 0.998. On the public MVTec 3D-AD benchmark, Cycle-CFM reaches a mean P-AUROC of 0.960 and improves accuracy by 37.5 % for elongated anomalies. With a runtime of 11.03 FPS and 469.52 MB of parameters, the model highlights both its effectiveness and deployability for real-time industrial inspection.}, archive = {J_ESWA}, author = {Yikang Shi and Xin Zhan and Yaqian Li and Zhongqiang Wu and Wenming Zhang and Haibin Li}, doi = {10.1016/j.eswa.2025.129745}, journal = {Expert Systems with Applications}, month = {3}, pages = {129745}, shortjournal = {Expert Syst. Appl.}, title = {Cycle-CFM: An unsupervised framework for robust multimodal anomaly detection in industrial settings}, volume = {298}, year = {2026}, }
• (2026). An explainable transfer learning approach to predict carbon emission intensity of coal-fired power plants with multi-source monitoring data. ESWA, 298, 129743. (www)

  @article{ , abstract = {Carbon mitigation policies and emission trading systems have heightened the need to monitor and predict the carbon emission intensity (CEI) of coal-fired power plants. Leveraging big data and machine learning (ML) technologies, this study trains predictive models of CEI using operational parameters, load rate, and coal-quality data from three Chinese power plants. The performance of Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Artificial Neural Network (ANN) was evaluated, and the challenge of limited data in individual plants was mitigated through instance-based transfer learning (ITL) and graft learning (GL) technologies. The results indicate that while traditional ML models struggle with poor data quality and limited samples, transfer learning between different plants will improve predictive accuracy substantially, and GL delivers the greatest gains. Among the most influential features, air supply temperature and load rate critically impact CEI and therefore should be carefully managed to achieve emission reductions. Nevertheless, the effectiveness of transfer learning depends on source data quality and model choice, and the proposed operational strategies require further validation before practical adoption. Our findings offer a robust framework for enhancing the predictive accuracy of plant CEI under data-scarce conditions and inform effective strategies for promoting a low-carbon transition in the energy sector.}, archive = {J_ESWA}, author = {Xiaodong Jin and Lingzhen Zhang and Fangyi Li and Wu Xie and Dawei Ma and Yan Wu}, doi = {10.1016/j.eswa.2025.129743}, journal = {Expert Systems with Applications}, month = {3}, pages = {129743}, shortjournal = {Expert Syst. Appl.}, title = {An explainable transfer learning approach to predict carbon emission intensity of coal-fired power plants with multi-source monitoring data}, volume = {298}, year = {2026}, }
• (2026). Uncertainty-guided denoising bi-classifier adversarial domain adaptation network for cross-domain fault diagnosis. ESWA, 298, 129742. (www)

  @article{ , abstract = {Intelligent fault diagnosis is crucial for ensuring the safety and reliability of modern industrial systems. However, the performance of deep learning models often significantly degrades due to the domain shift between training and testing data. Domain Adaptation (DA) methods, particularly bi-classifier adversarial networks, have proven effective in transferring knowledge from a labeled source domain to an unlabeled target domain. However, existing approaches often pay insufficient attention to target sample prediction accuracy, resulting in reduced feature discriminability and generalization. Additionally, due to the absence of labeled target data, most approaches rely on pseudo-labels, which are often noisy and unreliable, especially in the early stages of training. To address these issues, this paper proposes a novel uncertainty-guided denoising bi-classifier adversarial domain adaptation network (UGDBAN) for cross-domain fault diagnosis. Specifically, a feature generator based on Transformer layers is designed to capture long-range dependencies and local features. To mitigate the impact of noisy pseudo-labels, an uncertainty-based denoising pseudo-labeling mechanism is introduced to enhance the discriminability of features by redefining pseudo-labels and dynamically selecting high-confidence samples as clean samples. Building upon this denoised pseudo-label set, a Dirichlet uncertainty estimation-based class prototype alignment strategy is proposed to align domain features at the class level by selecting low-uncertainty samples representative of each class as prototypes. Extensive experiments demonstrate the effectiveness of UGDBAN, and comparative results with mainstream methods highlight its superiority.}, archive = {J_ESWA}, author = {Zheng Li and Lei Geng and Yanbei Liu and Feng Rong and Ming Ma and Jun Tong and Zhitao Xiao}, doi = {10.1016/j.eswa.2025.129742}, journal = {Expert Systems with Applications}, month = {3}, pages = {129742}, shortjournal = {Expert Syst. Appl.}, title = {Uncertainty-guided denoising bi-classifier adversarial domain adaptation network for cross-domain fault diagnosis}, volume = {298}, year = {2026}, }
• (2026). Partially view-aligned clustering via data recoupling and elastic bi-consistency learning. ESWA, 298, 129741. (www)

  @article{ , abstract = {Modern multi-view data often suffer from partial view alignment issues, yet most existing multi-view clustering (MVC) methods assume that the data are fully aligned, which is rarely the case in real-world scenarios. This misalignment leads to False Negative Pairs (FNPs), disrupting the learning process. While some methods address partial alignment, they often neglect intra-view consistency and multi-scale inter-view relationships, limiting their ability to capture both global and local structural dependencies. Additionally, the prevalent use of Mean Squared Error (MSE) as a reconstruction loss is suboptimal for discrete data, potentially causing severe performance degradation. To overcome these limitations, we propose Partially View-aligned Clustering via Data Recoupling and Elastic Bi-consistency Learning (PVC-DREBL). Our method integrates two key components: (1) a Data Recouple Module, which realigns the data to mitigate the effects of FNPs while leveraging an exponential contrastive loss to enhance learning stability and prevent overfitting; (2) an Elastic Bi-consistency Learning Module, designed to reconstruct diverse data types robustly while enforcing intra-view and multi-scale inter-view consistency. Extensive experiments on six benchmark datasets demonstrate that PVC-DREBL significantly outperforms existing methods, highlighting its effectiveness in handling partially view-aligned clustering tasks.}, archive = {J_ESWA}, author = {Wenzhe Liu and Jiongcheng Zhu and Jingbo Tan and Huibing Wang and Yong Zhang}, doi = {10.1016/j.eswa.2025.129741}, journal = {Expert Systems with Applications}, month = {3}, pages = {129741}, shortjournal = {Expert Syst. Appl.}, title = {Partially view-aligned clustering via data recoupling and elastic bi-consistency learning}, volume = {298}, year = {2026}, }
• (2026). Attention-based CNN for lithological identification from logging curves: A case study in the qaidam basin, china. ESWA, 298, 129740. (www)

  @article{ , abstract = {Subsurface lithological distribution is essential for extrapolating geological information from core to block or basin scales. Given the limited availability of core data, there is a critical need to develop a reliable method for establishing robust correlations between logging curves and lithologies in cores, thereby maximizing the value of large historical logging data. Here, we propose a novel attention-based convolutional neural network (ATT-CNN), which employs a 1D-CNN to transform six types of logging curves into high-dimensional feature space at each depth, and applies an attention mechanism to the 1D-CNN outputs along both the depth and feature dimensions. The architecture is designed to mimic human perceptual processing for lithology identification, leveraging curve combination, thresholding, and local pattern recognition within this enriched and high-dimensional feature representation. In addtion, the study employs wavelet-based preprocessing on logging curves to eliminate the impact of compaction-induced data drift on model generalization—an issue rarely considered in prior studies. The result showes that: ① The proposed ATT-CNN model demonstrates superior performance over benchmark models—the bidirectional gated recurrent unit (BiGRU) and an ensemble of machine learning models (En-ML)—across all evaluation metrics; ②Wavelet-based preprocessing enhances the generalization capability of both ATT-CNN and BiGRU, yielding higher metric scores and improved predictions, particularly in shallow-depth intervals; ③ For blind wells, the ATT-CNN outperforms BiGRU and En-ML in both accuracy and its ability to capture lithological variations even from low-amplitude curve deviations. The integration of ATT-CNN with wavelet-based preprocessing demonstrates significant potential for accurately characterizing subsurface lithological distribution, then provides critical support for key petroleum geology workflows, including provenance analysis, sedimentary facies mapping, and reservoir property prediction.}, archive = {J_ESWA}, author = {Jianguo Yin and Shuai Zhang and Zhixiong Wu and Shouji Pang and Rui Wang}, doi = {10.1016/j.eswa.2025.129740}, journal = {Expert Systems with Applications}, month = {3}, pages = {129740}, shortjournal = {Expert Syst. Appl.}, title = {Attention-based CNN for lithological identification from logging curves: A case study in the qaidam basin, china}, volume = {298}, year = {2026}, }
• (2026). Clustering-based brain functional segmentation via deep collapsed nonparametric von mises-fisher mixture models. ESWA, 298, 129739. (www)

  @article{ , abstract = {This paper presents a clustering-based framework for the analysis of functional magnetic resonance imaging (fMRI) data, with a particular focus on brain segmentation into functional sub-regions. The proposed approach comprises two key modules: representation learning and brain functional segmentation. To extract meaningful latent representations from high-dimensional fMRI signals while preserving temporal dependencies, we introduce the Spherical Variational Recurrent Autoencoder (SVRAE), a deep generative model built upon the Variational Autoencoder (VAE) architecture. Unlike conventional VAEs that assume a Gaussian prior, SVRAE employs the von Mises-Fisher (vMF) distribution to model latent variables on a unit hypersphere, which is more suitable for L 2 -normalized data. To further enhance temporal modeling, we replace standard fully connected layers with Long Short-Term Memory (LSTM) networks. For the segmentation module, we adopt a Collapsed Nonparametric von Mises-Fisher Mixture Model (Co-vMFMM), formulated within a Bayesian nonparametric framework. This model automatically adapts its complexity to the input data without requiring a predefined number of clusters. An efficient variational Bayes learning algorithm is developed to perform inference in a collapsed parameter space. Extensive experiments on publicly available fMRI datasets demonstrate the effectiveness and robustness of the proposed method in delineating functionally coherent brain sub-regions.}, archive = {J_ESWA}, author = {Wentao Fan and Wenchuan Zhang and Xiao Dong and Nizar Bouguila}, doi = {10.1016/j.eswa.2025.129739}, journal = {Expert Systems with Applications}, month = {3}, pages = {129739}, shortjournal = {Expert Syst. Appl.}, title = {Clustering-based brain functional segmentation via deep collapsed nonparametric von mises-fisher mixture models}, volume = {298}, year = {2026}, }
• (2026). An optimized hierarchical path planning method based on deep reinforcement learning for mobile robots. ESWA, 298, 129736. (www)

  @article{ , abstract = {With the development of elderly mobile robot technology, the training efficiency and path planning of robots have become key issues in research. Current mobile robot training faces challenges such as local optima and slow convergence speeds. To address these issues, this paper proposes a hierarchical path planning method based on deep reinforcement learning (H-DDQN). This method first introduces a global path planning module, which uses a density function to select high-value key points. In areas with dense obstacles, these key points provide the robot with effective global path guidance, enabling it to avoid getting stuck in a local optimum due to reliance solely on local information. In the local path planning module, we introduce a spatial attention mechanism based on global information, which weights global path points to enable the robot to focus on critical areas, thereby enhancing local decision-making capabilities and addressing dynamic obstacles on the basis of global optimality. Finally, this paper designs a new comprehensive reward function that combines path guidance from global key points with goal-oriented dense rewards, avoiding excessive unnecessary exploration and providing timely feedback to accelerate the model’s convergence speed. Experimental results show that compared to other existing algorithms, the H-DDQN algorithm converges more quickly during training and generates shorter and more efficient paths. In dynamic obstacle environments, the algorithm also demonstrates strong adaptability, combining global and local capabilities to achieve superior performance.}, archive = {J_ESWA}, author = {Jilin Zhang and Jia Qiao and Ke Huang and Menghua Zhang}, doi = {10.1016/j.eswa.2025.129736}, journal = {Expert Systems with Applications}, month = {3}, pages = {129736}, shortjournal = {Expert Syst. Appl.}, title = {An optimized hierarchical path planning method based on deep reinforcement learning for mobile robots}, volume = {298}, year = {2026}, }
• (2026). Assessing the renewable energy sources for sustainable energy generation systems: Interval-valued q-rung orthopair fuzzy SWARA-TOPSIS. ESWA, 298, 129735. (www)

  @article{ , abstract = {Renewable Energy Sources (RESs) help decarbonize power systems, but selecting among them is a challenging decision problem due to multiple, often conflicting, technical, economic, environmental, and health-related criteria. Consequently, numerous studies in the literature have attempted to address this decision-making issue using objective, subjective, and fuzzy decision-making procedures. However, there are still unaddressed research gaps in the literature, particularly regarding the explicit modeling of expert hesitation and ambiguity in real-world RES selection cases. The current study develops a decision-making model based on Step-wise Weight Assessment Ratio Analysis (SWARA) and Technique of Order Preference Similarity to the Ideal Solution (TOPSIS) methods integrated with Interval-Valued q-Rung Orthopair Fuzzy Sets (IV-q-ROFSs) to fill these gaps. Unlike previous studies that have predominantly applied conventional fuzzy MCDM techniques, our model introduces the first integration of IV-q-ROFS into RES selection. This novelty enables a more accurate representation of expert hesitation and uncertainty. The study is applied to a real industrial case in Turkey, where six RES alternatives are evaluated across 43 criteria by five senior experts under the supervision of a three-member professionals’ board. Furthermore, the structured robustness check and systematic literature mapping ensure that the proposed approach is methodologically robust and practically relevant for policymakers and energy planners. The application results of the developed model demonstrate that the estimated energy production potential of the RES and the effects of carcinogens generated from utilizing these energy sources are the critical factors influencing the selection of the most appropriate RESs. Solar energy ranked first among the alternatives. The applicability and validity of the developed model are examined by a comprehensive robustness check consisting of tests of sensitivity, comparison, and resilience to the rank reversal problem. Overall, the study provides (i) a novel methodological framework integrating IV-q-ROFS with SWARA and TOPSIS, (ii) empirical evidence from a comprehensive real-world RES selection case, and (iii) policy-relevant insights into the drivers of renewable energy adoption.}, archive = {J_ESWA}, author = {Ömer Faruk Görçün and Ahmet Aytekin and Selçuk Korucuk and Erfan Babaee Tirkolaee}, doi = {10.1016/j.eswa.2025.129735}, journal = {Expert Systems with Applications}, month = {3}, pages = {129735}, shortjournal = {Expert Syst. Appl.}, title = {Assessing the renewable energy sources for sustainable energy generation systems: Interval-valued q-rung orthopair fuzzy SWARA-TOPSIS}, volume = {298}, year = {2026}, }
• (2026). Enhancing liver cancer detection: An innovative deep learning approach combining GAN, ResNet, and vision transformer. ESWA, 298, 129734. (www)

  @article{ , abstract = {Liver cancer is a complex and life-threatening disease with significant diagnostic and therapeutic challenges. Automated liver cancer detection assists radiologists in identifying tumors and their severity accurately. In recent years, several deep-learning techniques have been implemented for diagnosing liver tumors and classification. Despite advancements in deep learning for medical imaging, existing liver cancer detection approaches continue to face several critical limitations. These include suboptimal diagnostic accuracy due to inadequate feature extraction, excessive computational demands that hinder real-time deployment, significant class imbalance within medical datasets leading to biased predictions, and overfitting caused by limited annotated training data. To address these challenges, this study introduces a novel and automated deep learning framework called CustomLiverNet, specifically designed for accurate and efficient liver cancer diagnosis using Computed Tomography images. The Generative Adversarial Network is introduced for generating realistic synthetic images, effectively improving the performance of the proposed technique and reducing class imbalance problems. The proposed technique integrates the strengths of Residual Networks and Vision Transformer to extract significant information from the input images and further enhance the performance of the proposed framework. The Residual Networks capture both low-level and high-level semantic features, whereas the Vision Transformer derives global and contextual feature representations from the input images. The model designs a customized fusion layer for combining the extracted features from both Residual Networks and Vision Transformer models. The classification layer predicts whether the liver tumor is benign or malignant. Further, Gradient-Weighted Class Activation Mapping is applied to highlight the critical regions of the image to enhance model transparency. The CustomLiverNet framework was trained and validated on two publicly available liver cancer datasets, including the liver tumor segmentation dataset, which contains 131 contrast-enhanced abdominal Computed Tomography scans, and the 3D image reconstruction for comparison of algorithm database, which includes 20 computed tomography scans. Experimental evaluation using standard metrics shows that CustomLiverNet achieved an accuracy of 98.79 %, precision of 98.64 %, recall of 98.58 %, and specificity of 98.35 %. These results demonstrate that the proposed model holds strong potential for enhancing early and accurate liver cancer diagnosis compared to previous studies.}, archive = {J_ESWA}, author = {Shivani Joshi and Avinash Dwivedi and Rajiv Kumar and Ashish Kumar and Raju Kumar and Amrita}, doi = {10.1016/j.eswa.2025.129734}, journal = {Expert Systems with Applications}, month = {3}, pages = {129734}, shortjournal = {Expert Syst. Appl.}, title = {Enhancing liver cancer detection: An innovative deep learning approach combining GAN, ResNet, and vision transformer}, volume = {298}, year = {2026}, }
• (2026). Unleashing the power of large language models for low-resource relation triplet extraction by structure-to-text data generation. ESWA, 298, 129733. (www)

  @article{ , abstract = {Scaling language models have revolutionized widespread NLP tasks, yet little investigation has been conducted to assess the ability of Large Language Models (LLMs) to explore low-resource relation triplet extraction. This paper investigates essential methodologies, k -shot demonstration of in-context learning, and many-shot instruction tuning for few-shot and zero-shot relation triplet extraction using FlanT5, supported by exhaustive experiments. To enhance low-resource setting performance, we further propose different types of demonstration examples and task-related instructions for data generation. Specifically, we leverage LLMs to construct a structured prompt template for generating synthetic training data based on structured text and efficiently explore the boundary issues of examples in both instruction tuning and in-context learning, assessing their impact on model performance. To address the challenge of extracting multiple relation triplets from a single sentence, we design a novel Multiple Triplet Search (MTS) algorithm. Furthermore, we find that in-context learning can match the performance of previous prompt learning methods. Additionally, integrating synthetic data with the LLM can improve existing solutions in low-resource scenarios, achieving new state-of-the-art results. Experiments conducted on six relation extraction datasets demonstrate the efficacy of the proposed model for the zero-shot and few-shot RTE tasks. Our source code is publicly available at https://github.com/Phevos75/LLMRTE.}, archive = {J_ESWA}, author = {Qian Guo and Yi Guo and Jin Zhao}, doi = {10.1016/j.eswa.2025.129733}, journal = {Expert Systems with Applications}, month = {3}, pages = {129733}, shortjournal = {Expert Syst. Appl.}, title = {Unleashing the power of large language models for low-resource relation triplet extraction by structure-to-text data generation}, volume = {298}, year = {2026}, }
• (2026). LDATA-net: Dynamic feature adaptation for efficient feature learning in resource-limited UAV detection. ESWA, 298, 129725. (www)

  @article{ , abstract = {Unmanned Aerial Vehicle (UAV) image analysis faces the dual challenges of complex background interference and limited onboard computational resources, particularly when processing extreme scale variations across multiple viewpoints. Existing approaches typically enhance detection accuracy by increasing model complexity, but this often leads to parameter proliferation that exceeds the deployment limits of airborne platforms. To address this fundamental contradiction, we propose LDATA-Net (Lightweight Dynamic Aggregation Task-Aligned Network), which pioneers a “Dynamic Feature Adaptation” design paradigm aimed at achieving synergistic optimization between parameter efficiency and detection accuracy. This framework systematically realizes end-to-end dynamic adaptive capabilities through three core components that operate collaboratively across feature extraction, fusion, and detection stages: (1) Dynamic Multi-Branch Depthwise Block (DMBD-Block), whose core innovation is our proposed novel operator DIDWConv, which adaptively adjusts receptive fields according to input features to capture targets of extreme scales and orientations; (2) Lightweight Dynamic Aggregation Network (LDANet), which effectively preserves critical spatial contextual information through hierarchical fusion architecture and dynamic weighting mechanisms; (3) Dynamic Adaptive Head (DA-Head), which effectively mitigates task conflicts through geometric and semantic dynamic feature alignment. LDATA-Net achieves 35.4 %, 77.9 %, and 51.2 % AP 50 on VisDrone2019, DOTA1.0, and AI-TODv2 datasets respectively with only 2.8M parameters, establishing a new paradigm for designing memory-efficient yet high-performance detection systems, particularly for resource-constrained heterogeneous computing aviation platforms.}, archive = {J_ESWA}, author = {Shuming Lin and Sang Feng and Junnan Tan}, doi = {10.1016/j.eswa.2025.129725}, journal = {Expert Systems with Applications}, month = {3}, pages = {129725}, shortjournal = {Expert Syst. Appl.}, title = {LDATA-net: Dynamic feature adaptation for efficient feature learning in resource-limited UAV detection}, volume = {298}, year = {2026}, }
• (2026). Exploring intra- and inter-organizational collaboration opportunities across a technological knowledge ecosystem: A metapath2vec approach. ESWA, 298, 129724. (www)

  @article{ , abstract = {Due to the increasing complexity and scale of technological knowledge ecosystems, organizations face challenges in identifying intra- and inter-organizational collaboration opportunities. In this respect, prior studies have proposed patent-based approaches, but they are subject to several limitations: (1) insufficient consideration of technological relationships within the ecosystem, (2) simplified unit of analysis, and (3) limited organization-centric assessments. This study proposes a network embedding and text-reranking approach to explore potential intra- and inter-organizational collaboration opportunities. First, the technological knowledge ecosystem is represented as a heterogeneous patent network comprising patents, inventors, assignees, and technology classification codes. Second, inventor nodes are embedded using metapath2vec, which performs random walks along predefined metapaths to capture diverse knowledge flows within the ecosystem. Third, potential collaborators are explored through (1) screening candidates based on technological reachability, which measures the possibility of knowledge exploration based on contextual similarity within the network, and (2) reranking candidates based on technological relevance, which quantifies the possibility of knowledge exploitation based on the similarity of technological know-how and experiences. Finally, ten quantitative patent indicators are developed to assess the implications of these opportunities at both the inventor and organization levels. The validity of the proposed approach is demonstrated through a case study involving 28,888 US patents and 9,196 inventors in the field of energy storage technology. This study contributes to advancing the theoretical understanding of technological knowledge ecosystems while also serving as a supplementary tool to explore organizational collaboration opportunities.}, archive = {J_ESWA}, author = {Jaemin Chung and Jaewoong Choi and Janghyeok Yoon}, doi = {10.1016/j.eswa.2025.129724}, journal = {Expert Systems with Applications}, month = {3}, pages = {129724}, shortjournal = {Expert Syst. Appl.}, title = {Exploring intra- and inter-organizational collaboration opportunities across a technological knowledge ecosystem: A metapath2vec approach}, volume = {298}, year = {2026}, }
• (2026). An improved actor-critic architecture with PPO for the traveling salesman problem. ESWA, 298, 129723. (www)

  @article{ , abstract = {The traveling salesman problem (TSP) is a classic NP-hard problem in combinatorial optimization with extensive practical applications. In this paper, we present an improved Actor-Critic architecture incorporating Proximal Policy Optimization (PPO) to effectively solve TSP. We introduce adaptive temperature scheduling, comprehensive state representation, and layer normalization to enhance learning stability. Experimental results demonstrate our Improved Actor-Critic approach achieves significant improvements ranging from 8.7 % to 55.9 % for different problem sizes compared to established reinforcement learning baselines including Q-Learning, SARSA, Double Q-Learning, Actor-Critic with Experience Replay (ACER), and Trust Region Policy Optimization (TRPO), with particularly strong performance on smaller instances between 20 to 100 cities. When testing on standard TSPLIB benchmarks, our method shows consistent advantages of 12 % to 33 % compared to classical approaches While tabular methods become computationally infeasible beyond 250 cities due to memory constraints, our approach maintains high solution quality for problems up to 1432 cities on our experimental setup (Intel® Core™i9-10900X CPU @ 3.70GHz × 20 with four NVIDIA Quadro RTX 5000 GPUs). Our ablation studies confirm the importance of each component in our proposed architecture, in which the improved state representation provides the most significant contribution to our model performance. This research significantly advances reinforcement learning approaches to combinatorial optimization, with practical implications for logistics, telecommunications, and manufacturing. The developed source code is available at: https://github.com/LetuQingge/TSP_Environment .}, archive = {J_ESWA}, author = {Hailemicael Lulseged Yimer and Pei Yang and Letu Qingge}, doi = {10.1016/j.eswa.2025.129723}, journal = {Expert Systems with Applications}, month = {3}, pages = {129723}, shortjournal = {Expert Syst. Appl.}, title = {An improved actor-critic architecture with PPO for the traveling salesman problem}, volume = {298}, year = {2026}, }
• (2026). Synchronization of semi-markov jump two-time-scale fuzzy neural networks dealing with dual-scale hybrid attacks and its application. ESWA, 298, 129718. (www)

  @article{ , abstract = {In this paper, the synchronization issue of the semi-Markov jump two-time-scale fuzzy neural networks dealing with dual-scale hybrid attacks is addressed, in which the dual-scale hybrid attacks mean that hybrid attacks can be encountered in different time scales transmission channels. Based on mismatched membership functions, a new ϵ -dependent combined synchronization controller is proposed to ensure the synchronization of the master semi-Markov jump two-time-scale fuzzy neural networks and the slave ones while it is capable of resisting independent dual-scale hybrid attacks. Through the development of a Lyapunov function incorporating the singular perturbation parameter ϵ , stability conditions and a computational approach for determining the synchronization controller gain are derived for semi-Markov jump two-time-scale fuzzy neural networks. Finally, some simulations and two encryption and decryption processes are used to show the effectiveness of obtained results.}, archive = {J_ESWA}, author = {Feng Li and Ya-Nan Wang and Lei Su and Sangmoon Lee}, doi = {10.1016/j.eswa.2025.129718}, journal = {Expert Systems with Applications}, month = {3}, pages = {129718}, shortjournal = {Expert Syst. Appl.}, title = {Synchronization of semi-markov jump two-time-scale fuzzy neural networks dealing with dual-scale hybrid attacks and its application}, volume = {298}, year = {2026}, }
• (2026). Automated recognition and measurement of cellular and dendritic microstructures in alloys via machine learning. ESWA, 298, 129717. (www)

  @article{ , abstract = {The dendritic or cellular morphologies of alloys and metals formed during casting processes significantly influence key properties such as mechanical strength, toughness, hardness, and electrical or thermal conductivities. The literature presents correlations between these properties and the microstructural length scale, which requires accurate characterization at the micrometer level. However, traditional evaluation methods demand extensive experimental efforts, including careful metallographic preparation, high-quality imaging, and a large number of measurements for statistical reliability - often relying on analyst proficiency. This study proposes a machine learning-based workflow tailored for the automated processing of microstructure images. The approach enables the autonomous measurement of key microstructural features while minimizing bias and inconsistencies among analysts. By integrating advanced image processing techniques with object detection algorithms based on Convolutional Neural Networks (CNNs), the method autonomously identifies microstructural morphologies and quantifies their spacing scales. Three model types-Cell, Dendrite, and Hybrid (exhibiting both dendritic and cellular features)-were trained and validated on using a dataset of 200 images. Among them, the Cell detection model achieved the highest performance, with a mean Average Precision (mAP) of 78.77 %, followed by the Hybrid (75.63 %) and Dendrite (72.87 %) models. Finally, the automated measurements models were applied to literature images and compared to reported microstructural growth correlations.}, archive = {J_ESWA}, author = {Guilherme Marim da Silva and Rafael Kakitani and Carlos Henrique da Silva Santos and Amauri Garcia and Noé Cheung}, doi = {10.1016/j.eswa.2025.129717}, journal = {Expert Systems with Applications}, month = {3}, pages = {129717}, shortjournal = {Expert Syst. Appl.}, title = {Automated recognition and measurement of cellular and dendritic microstructures in alloys via machine learning}, volume = {298}, year = {2026}, }
• (2026). Online reinforcement learning strategies driven artificial bee colony algorithm for bi-objective distributed reentrant flowshops scheduling problem with sequence-sependent setup time. ESWA, 298, 129716. (www)

  @article{ , abstract = {With the increasing complexity of manufacturing systems, reentrancy has become prevalent in many production environments. This study investigates a bi-objective distributed reentrant flow shop scheduling problem with sequence-dependent setup times (DRFSP-SDST). The objectives are to minimize the total energy consumption (TEC) and the maximum completion time (makespan), simultaneously. First, a bi-objective mathematical model for the DRFSP-SDST is formulated based on practical reentrant production scenarios. Second, the artificial bee colony (ABC) algorithm and its variants are employed to solve the DRFSP-SDST. According to the characteristics of the DRFSP-SDST, six local search operators are specifically designed to enhance the performance of the proposed algorithms. For promoting greener and more energy-efficient production, two speed-scaling strategies are developed. Third, two reinforcement learning (RL) algorithms, Q-learning and State-Action-Reward-State-Action (SARSA), are integrated into the iterative process as online learning strategies to guide the selection of high-quality local search strategies during the iterations of the proposed algorithms. For each RL algorithm, two distinct selection strategies for local search operators are designed. Finally, the effectiveness of the proposed enhancement strategies is evaluated through comprehensive numerical experiments on 36 benchmark instances. The performance of the proposed algorithms is further validated via the Friedman test. The experimental results and analysis demonstrate that the ABC algorithm enhanced by SARSA-based local search exhibits superior competitiveness in solving the DRFSP-SDST.}, archive = {J_ESWA}, author = {Ao Yao and Kaizhou Gao and Ponnuthurai Nagaratnam Suganthan and Hongyan Sang}, doi = {10.1016/j.eswa.2025.129716}, journal = {Expert Systems with Applications}, month = {3}, pages = {129716}, shortjournal = {Expert Syst. Appl.}, title = {Online reinforcement learning strategies driven artificial bee colony algorithm for bi-objective distributed reentrant flowshops scheduling problem with sequence-sependent setup time}, volume = {298}, year = {2026}, }
• (2026). Defects inspection system for building facades using drones and deep learning method. ESWA, 298, 129715. (www)

  @article{ , abstract = {Regular inspection and maintenance of building facades are essential for preserving structural integrity and aesthetic quality, especially in aging urban high-rises. While drone-based visual inspection powered by artificial intelligence (AI) offers benefits in speed, safety, and scalability, existing methods are typically limited to single defect types or uniform facade categories due to the challenges of detecting multi-scale defects in complex, heterogeneous environments. This study introduces an automated multiclass defects inspection system for building facades by integrating drone technology, an AI-driven segmentation platform, and automatic report generation. Central to the system is a segmentation AI model capable of detecting multiclass defects with orders-of-magnitude differences in scale across diverse facade backgrounds. To handle the pixel imbalance of defects ranging from fine cracks to large spalling and glass breakage, the model is built upon EfficientUNet++, trained on a carefully curated dataset and optimized using adjustable batch sizes and active learning rates to improve multi-scale feature learning and mitigate overfitting. Evaluations on validation and out-of-sample datasets demonstrate that the proposed model achieves superior performance across all defect classes. Real-world drone experiments further confirm the model’s practical applicability, with high recall rates in detecting spalling, water seepage, cracks, and glass breakage across different types of facades. This work pioneers a robust, scalable, and efficient AI-based framework for automated multiclass facade defect inspection, providing actionable information for engineers and supporting urban infrastructure maintenance.}, archive = {J_ESWA}, author = {Xiaoling Zhou and Robert Lee Kong Tiong}, doi = {10.1016/j.eswa.2025.129715}, journal = {Expert Systems with Applications}, month = {3}, pages = {129715}, shortjournal = {Expert Syst. Appl.}, title = {Defects inspection system for building facades using drones and deep learning method}, volume = {298}, year = {2026}, }
• (2026). Cross-modal multitask learning for automated quantitative characterization of infrastructure airhole defects. ESWA, 298, 129713. (www)

  @article{ , abstract = {Quantitative characterization of apparent quality defects in infrastructure is a crucial component of operations and maintenance. It enables rapid assessment of defect severity and supports the timely formulation of preventive strategies. However, a singular visual modality struggles to simultaneously ensure the dual tasks of defect detection and measurement accuracy. To solve these problems, this paper proposes a novel framework for cross-modal multitask learning networks, which comprehensively integrates the advantages of image detection and point cloud measurement. The pixel points identified in the image are mapped to their corresponding three-dimensional coordinates in the point cloud through intensive feature matching. A measurement strategy for the inherent characteristics of the defect is subsequently proposed. Based on prior knowledge of the defect, the area and volume of defects are quantified accurately. Finally, extensive experiments on detection, matching and measurement demonstrate the efficacy of the proposed method. The results provide a valuable reference for the quantitative characterization of infrastructure defects.}, archive = {J_ESWA}, author = {Yu Wang and Yingchao Dai and Xiaodong Gan and Zhengtao Yang and Zhou Wu}, doi = {10.1016/j.eswa.2025.129713}, journal = {Expert Systems with Applications}, month = {3}, pages = {129713}, shortjournal = {Expert Syst. Appl.}, title = {Cross-modal multitask learning for automated quantitative characterization of infrastructure airhole defects}, volume = {298}, year = {2026}, }
• (2026). Language proficiency assessment of autistic children using large language models. ESWA, 298, 129712. (www)

  @article{ , abstract = {Language impairment is a common comorbidity in children with autism spectrum disorder (ASD), and language proficiency assessment is a primary method for identifying such impairments. However, traditional assessment tools are often subjective and inefficient, while existing computer-assisted methods are limited by a narrow focus and insufficient use of natural language samples. To address these issues, this study proposes a framework for assessing children’s language abilities based on large language models (LLMs). We first preprocess the natural language samples from children and design multiple assessment dimensions and workflows. To enhance the stability of the assessment, we introduce a multi-expert voting mechanism and perform a comparative analysis of various large language models’ performance. The experimental results demonstrate a strong correlation between the framework’s assessment results and the Mullen Scales of Early Learning (MSEL) verbal developmental quotients, with a Pearson correlation coefficient of 0.8 ( p < 0.001). Furthermore, the results show that the multi-dimensional evaluation can accurately differentiate between ASD and typically developing (TD) children, achieving a classification accuracy of 0.98. These findings suggest that the proposed framework has significant potential for improving the accuracy of ASD identification.}, archive = {J_ESWA}, author = {Saige Qin and Min Liu and Tongquan Wei and Qiaoyun Liu}, doi = {10.1016/j.eswa.2025.129712}, journal = {Expert Systems with Applications}, month = {3}, pages = {129712}, shortjournal = {Expert Syst. Appl.}, title = {Language proficiency assessment of autistic children using large language models}, volume = {298}, year = {2026}, }
• (2026). EQUINAS: Equilibrium-guided differentiable neural architecture search. ESWA, 298, 129711. (www)

  @article{ , abstract = {Recent research has significantly mitigated the performance collapse issue in Differentiable Architecture Search (DARTS) by either refining architecture parameters to better reflect the true strengths of operations or developing alternative metrics for evaluating operation significance. However, the actual role and impact of architecture parameters remain insufficiently explored, creating critical ambiguities in the search process. To address this gap, we conduct a rigorous theoretical analysis demonstrating that the change rate of architecture parameters reflects the sensitivity of the supernet’s validation loss in architecture space, thereby influencing the derived architecture’s performance by shaping supernet training dynamics. Building on these insights, we introduce the concept of a Stable Equilibrium State to capture the stability of the bi-level optimization process and propose the Equilibrium Influential ( E I ) metric to assess operation importance. By integrating these elements, we propose EQUINAS, a differentiable NAS approach that leverages the Stable Equilibrium State to identify the optimal state during the search process and derives the final architecture using the E I metric. Extensive experiments across diverse datasets and search spaces demonstrate that EQUINAS achieves competitive test accuracy compared to state-of-the-art methods while significantly reducing search costs. Additionally, EQUINAS shows remarkable performance in Transformer-based architectures and excels in real-world applications such as image classification and text recognition.}, archive = {J_ESWA}, author = {Weisheng Xie and Xiangxiang Gao and Xuwei Fang and Hui Li and Chen Hang and Shaoyuan Li}, doi = {10.1016/j.eswa.2025.129711}, journal = {Expert Systems with Applications}, month = {3}, pages = {129711}, shortjournal = {Expert Syst. Appl.}, title = {EQUINAS: Equilibrium-guided differentiable neural architecture search}, volume = {298}, year = {2026}, }
• (2026). Freq-DETR: Frequency-aware transformer for real-time small object detection in unmanned aerial vehicle imagery. ESWA, 298, 129710. (www)

  @article{ , abstract = {Recent advancements in unmanned aerial vehicle (UAV) and remote sensing technologies have propelled UAV object detection to the forefront of computer vision research. Despite significant progress in deep learning-based detection algorithms, critical challenges persist in small object detection, including high-frequency information loss, inadequate multiscale feature representation, etc. To address these limitations, this paper proposes Freq-DETR, a frequency-aware real-time transformer detection framework leveraging frequency domain analysis to enhance edge detail preservation and global contextual modeling through three novel innovations. First, the frequency-enhanced convolution module (FECM) synergistically integrates spatial and frequency features via dual-branch processing; Second, the decoupled intra-feature scale interaction module (DSC-Clo block) facilitates the integration of high-frequency local and low-frequency global information; Finally, the attention-guided selective feature pyramid network (AGS-FPN) employs context-aware attention for high-level screening feature fusion. Extensive evaluations on the VisDrone2019 benchmark demonstrate that Freq-DETR outperforms the baseline RT-DETR by 4.9 % m a p @ 50 gain while maintaining computational efficiency. There are also remarkable improvements on both UAVDT and HIT-UAV datasets. Ablation investigations and visual interpretability analyses further confirm the complementary benefits of its frequency-domain components and the framework’s robustness in complex aerial scenarios.}, archive = {J_ESWA}, author = {Jiayi Chen and Ningzhong Liu and Han Sun and Yu Wang}, doi = {10.1016/j.eswa.2025.129710}, journal = {Expert Systems with Applications}, month = {3}, pages = {129710}, shortjournal = {Expert Syst. Appl.}, title = {Freq-DETR: Frequency-aware transformer for real-time small object detection in unmanned aerial vehicle imagery}, volume = {298}, year = {2026}, }
• (2026). Companion learning networks: A deep reinforcement learning algorithm with partner networks. ESWA, 298, 129709. (www)

  @article{ , abstract = {Existing deep reinforcement learning (DRL) agents suffer from severe reward instability during late-stage exploration, particularly when encountering novel states in complex continuous environments. A variety of existing studies focus on improving an agent’s reward exploration. However, they ignore the instability problem that arises when the agent faces new states in the later stages of exploration. This paper proposes a novel companion learning network (CLN) based on the idea that the guidance can accelerate human learning efficiency and reduce the risk of making mistakes. The CLN integrates a short-term partner network to intensively learn localized environmental patterns, offering adaptive action guidance for recent states. Simultaneously, a global Q-network dynamically incorporates the partner’s decaying guidance signals, balancing autonomous exploration with error mitigation. As training progresses, the partner’s influence gradually diminishes, allowing the Q-network to solidify robust policies without persistent dependence. Extensive experiments on four OpenAI Gym environments demonstrate that the CLN can significantly improve the exploration stability in most tested scenarios, achieving up to 49% reduction in late-stage reward standard deviation compared to baseline DRL methods.}, archive = {J_ESWA}, author = {Jin Xu and Jinfeng Bu and Yu Zhang and Jia-Dong Zhang and Chi-Yin Chow}, doi = {10.1016/j.eswa.2025.129709}, journal = {Expert Systems with Applications}, month = {3}, pages = {129709}, shortjournal = {Expert Syst. Appl.}, title = {Companion learning networks: A deep reinforcement learning algorithm with partner networks}, volume = {298}, year = {2026}, }
• (2026). Expensive multiobjective immune algorithm using a novel differential evolution in objective space. ESWA, 298, 129708. (www)

  @article{ , abstract = {Generating offspring solutions with strong convergence and diversity is critical when solving expensive multiobjective optimization problems due to the limited number of objective function evaluations. However, existing algorithms produce offspring solutions in the decision space, causing significant uncertainty in obtaining offspring with strong convergence and diversity. To address this issue, we devise a novel differential evolution based on the objective space rather than the decision space, called Differential Objective Evolution (DOE). Specifically, DOE generates objective values with strong convergence and diversity and then maps these values to the decision space to achieve high-quality offspring. Furthermore, we utilize a multiobjective immune algorithm to produce high-quality samples for effectively training the mapping in DOE. When compared with eleven recently proposed algorithms on 105 expensive multiobjective optimization problems, the experiments demonstrate the superiority of our algorithm and the contributions of DOE in both population convergence and diversity.}, archive = {J_ESWA}, author = {Yuchao Su and Wu Lin and Daxin Zhu and Anhui Tan and Ka-Chun Wong and Qiuzhen Lin}, doi = {10.1016/j.eswa.2025.129708}, journal = {Expert Systems with Applications}, month = {3}, pages = {129708}, shortjournal = {Expert Syst. Appl.}, title = {Expensive multiobjective immune algorithm using a novel differential evolution in objective space}, volume = {298}, year = {2026}, }
• (2026). Talk with graph: A fuzzy sentiment-aware framework for explainable recommendation. ESWA, 298, 129707. (www)

  @article{ , abstract = {User reviews reflect user preferences and item characteristics, optimizing the predictive accuracy and explanation generation of personalized recommendation systems. However, existing models face challenges due to subjective uncertainty in user feedback and a lack of transparency. Reviews often contain ambiguous emotional expressions, with the same product receiving positive, neutral, and negative sentiments. Many recommendation models assume alignment between ratings and review sentiments, but in practice, users may give high ratings while expressing dissatisfaction or vice versa. These inconsistencies complicate accurate modeling of user preferences. To address these issues, a Large Language Model (LLM)-driven sentiment-enhanced heterogeneous graph neural network framework is proposed. This framework jointly models interaction data and fuzzy sentiment information from reviews to improve both recommendation accuracy and explainability. By leveraging LLM with dual-prompt strategies, high-quality sentiment distributions and semantic insights are extracted. Review sentiments are then quantified using intuitionistic fuzzy numbers to address data sparsity and uncertainty, capturing implicit relationships between users, items, and entities in a sentiment-enhanced heterogeneous relational graph. A fuzzy sentiment-weighted graph convolutional network (FSGCN) is introduced for dynamic higher-order feature learning, adjusting sentiment weights based on user-item interactions and emotional context. The framework also integrates LLM-driven query interpretation to generate recommendations with transparent, context-aware rationales. This approach enables users to understand the reasoning behind recommendations, significantly enhancing explainability and trust.}, archive = {J_ESWA}, author = {Zhinan Li and Zhenyu Liu and Guodong Sa and Mingjie Hou and Jiacheng Sun and Jianrong Tan}, doi = {10.1016/j.eswa.2025.129707}, journal = {Expert Systems with Applications}, month = {3}, pages = {129707}, shortjournal = {Expert Syst. Appl.}, title = {Talk with graph: A fuzzy sentiment-aware framework for explainable recommendation}, volume = {298}, year = {2026}, }
• (2026). Geodesic-based path planning for port transfer robots on riemannian manifolds. ESWA, 298, 129706. (www)

  @article{ , abstract = {With the rapid intelligent transformation of the automotive industry and the surge in production volume, intelligent autonomous robots equipped with integrated perception and planning systems are playing an increasingly vital role in vehicle transfer operations. Optimizing dispatch paths of robots is essential for improving overall operational efficiency, yet achieving a balance among path length, feasibility, and safety margin remains a significant challenge. To address this issue, we propose a geodesic-based path planning method formulated on Riemannian manifolds. The approach jointly considers directional motion constraints, steering effort, and obstacle accessibility boundaries to construct a Riemannian metric tensor that encodes local path cost structures. This transforms the planning task into a geodesic shortest path problem, which is efficiently solved using the Geometric heat flow (GHF) method. The resulting paths naturally comply with kinematic constraints and exhibit strong obstacle-avoidance capabilities, significantly enhancing safety and executability. Extensive simulations and real-world experiments in high-density port yard environments demonstrate the practicality and robustness of the proposed method under complex spatial constraints and obstacle configurations.}, archive = {J_ESWA}, author = {Runjiao Bao and Junzheng Wang and Shoukun Wang}, doi = {10.1016/j.eswa.2025.129706}, journal = {Expert Systems with Applications}, month = {3}, pages = {129706}, shortjournal = {Expert Syst. Appl.}, title = {Geodesic-based path planning for port transfer robots on riemannian manifolds}, volume = {298}, year = {2026}, }
• (2026). A social balance theory-based modeling framework for group-to-empirical decision-making transition with cognitive inertia and trust propagation. ESWA, 298, 129705. (www)

  @article{ , abstract = {Inspired by empirical decision-making (EDM) processes, we propose a novel modeling framework where agents iteratively integrate social neighbors’ opinions into their cognitive inertia sequences (CISs), gradually prioritizing their accumulated CISs over time. This framework simulates the transition from group decision-making (GDM) to EDM through dynamic trust/distrust propagation and aggregation mechanisms grounded in social balance theory–capturing relational scenarios such as “a friend of a friend is a friend”, “a friend of an enemy is an enemy”, “an enemy of a friend is an enemy”, and “an enemy of an enemy is a stranger”. The paradigm incorporates two core mechanisms: (1) an endogenous cognitive inertia mechanism that uses the psychological serial-positioning effect to model cognitive inertia weights, accounting for primacy, recency, and U-shaped memory effects; and (2) an exogenous mechanism that quantifies comprehensive trust/distrust degrees via opinion similarity, stability similarity, and network structure similarity. To prevent followers from falling into cognitive freezing, a cluster leader-based consensus-reaching strategy is introduced. Extensive comparative experiments on real-world network datasets confirm the model’s effectiveness and robustness.}, archive = {J_ESWA}, author = {Jianglin Dong and Yiyi Zhao and Shangqun Mu and Haixia Mao and Jiangping Hu}, doi = {10.1016/j.eswa.2025.129705}, journal = {Expert Systems with Applications}, month = {3}, pages = {129705}, shortjournal = {Expert Syst. Appl.}, title = {A social balance theory-based modeling framework for group-to-empirical decision-making transition with cognitive inertia and trust propagation}, volume = {298}, year = {2026}, }
• (2026). OTPS-VO: Enhanced RGB-D odometry for indoor service robots leveraging structural features. ESWA, 298, 129704. (www)

  @article{ , abstract = {Simultaneous Localization and Mapping (SLAM) is essential for indoor service robots to achieve reliable navigation and mapping. While point and line features have been extensively utilized to enhance the accuracy of visual odometry (VO), current methods often overlook the rich geometric information embedded in the spatial relationships among structural lines. In particular, the parallelism and collinearity within groups of line segments are underexploited, and geometric constraints are typically applied only heuristically or post hoc, limiting robustness in low-texture and repetitive environments. To address these challenges, a robust VO system is proposed that integrates structural feature grouping with adaptive MW tracking. A unified feature extraction strategy is introduced to detect point and line features simultaneously, improving computational efficiency. To mitigate pose drift caused by unreliable line segments, a set of parallel line features is constructed based on local geometric constraints, and a novel reprojection error model is formulated to enhance pose estimation. Furthermore, a tracking strategy based on local Manhattan World (MW) structure is developed to ensure low-drift estimation across various structured indoor scenes. Extensive experiments on multiple public datasets and a custom-built service robot platform demonstrate that the proposed method outperforms existing state-of-the-art approaches under dynamic lighting conditions and in environments rich in lines and planes. The system also operates at a real-time speed of 30 frames per second, meeting the requirements of practical robotic applications.}, archive = {J_ESWA}, author = {Zhiyu Wang and Weili Ding and Ying Zhang and Changchun Hua}, doi = {10.1016/j.eswa.2025.129704}, journal = {Expert Systems with Applications}, month = {3}, pages = {129704}, shortjournal = {Expert Syst. Appl.}, title = {OTPS-VO: Enhanced RGB-D odometry for indoor service robots leveraging structural features}, volume = {298}, year = {2026}, }
• (2026). Adaptive gated meta graph retention network: A model for urban traffic flow prediction. ESWA, 298, 129703. (www)

  @article{ , abstract = {Traffic flow prediction is crucial for urban transportation systems. Existing models are still deficient in training efficiency and modeling dynamic spatial-temporal dependencies, various external factors, time-varying topology. Based on this, this paper introduces the Adaptive Gated Meta Graph Retention Network (AGMGRN), a novel model for spatial-temporal traffic flow prediction. Specifically, the AGMGRN integrates the attention mechanism with the retention network to model spatial-temporal dependencies. The AGMGRN proposes a gated dynamic connection block to enhance the model’s dynamic modeling capabilities. The AGMGRN considers the influence of external factors on traffic conditions through meta-learning approaches. The AGMGRN proposes an adaptive graph block to construct time-varying topologies. Experiments on four actual large-scale datasets demonstrate that the AGMGRN achieves superior prediction accuracy and high applicability.}, archive = {J_ESWA}, author = {Xing Li and Yuequan Bao}, doi = {10.1016/j.eswa.2025.129703}, journal = {Expert Systems with Applications}, month = {3}, pages = {129703}, shortjournal = {Expert Syst. Appl.}, title = {Adaptive gated meta graph retention network: A model for urban traffic flow prediction}, volume = {298}, year = {2026}, }
• (2026). DASF-GRL: Dynamic agent-scaling framework with game-augmented reinforcement learning for defensive counter-air operations. ESWA, 298, 129702. (www)

  @article{ , abstract = {Defensive Counter-Air (DCA) operations are pivotal for modern air defense, but existing studies are limited by static defender populations and oversimplified attacker models. We address these limitations with the Dynamic Agent-Scaling Framework with Game-Augmented Reinforcement Learning (DASF-GRL), which dynamically scales defender populations based on real-time threat levels. Specifically, we introduce a hybrid imitation-reinforcement training strategy that integrates attention mechanisms into critic networks to enable dynamic agent scaling. By incorporating a safety barrier function rooted in differential game theory, we constrain agents’ action spaces and enhance policy reliability. Furthermore, we developed a DCA simulation platform supporting reinforcement learning validation and designed a novel Apollonius-based penetration strategy for attackers to improve algorithmic robustness. Experiments demonstrate that DASF-GRL adaptively adjusts defender populations across scenarios involving 20, 40, and 60 attackers, markedly outperforming baseline methods in convergence speed and defense success rates. This framework offers novel theoretical paradigms and practical tools for intelligent decision-making in DCA environments.}, archive = {J_ESWA}, author = {Yuxuan Chen and He Luo and Guoqiang Wang}, doi = {10.1016/j.eswa.2025.129702}, journal = {Expert Systems with Applications}, month = {3}, pages = {129702}, shortjournal = {Expert Syst. Appl.}, title = {DASF-GRL: Dynamic agent-scaling framework with game-augmented reinforcement learning for defensive counter-air operations}, volume = {298}, year = {2026}, }
• (2026). Coordinated low-carbon economic scheduling of integrated electricity-heat-gas-hydrogen-methanol multi-microgrids considering electricity-methanol transaction. ESWA, 298, 129701. (www)

  @article{ , abstract = {The uneven spatial and temporal distribution of renewable energy resources poses significant challenges for multi-microgrid (MG) systems, resulting in high operational costs and low renewable energy utilization. To overcome these challenges, this work investigates a peer-to-peer electricity transaction and hydrogen-methanol-hydrogen technology-based methanol transaction among multi-MG. Besides, to realize net-zero emissions and carbon cycle utilization, the carbon capture system and hydrogen blending system are introduced into MG to reduce carbon dioxide emissions and capture and reform carbon dioxide for methanol synthesis equipment. Additionally, a cooperative operation model based on the Nash bargaining theory for multi-MGs under the transaction amount and price constraints of electricity and methanol is constructed. Due to the characteristics of non-convex and non-linear, the Nash bargaining is transformed into minimizing operation costs (sub-problem one) and maximizing payment benefits (sub-problem two). During the process of benefit allocation in sub-problem two, this work adopts a nonlinear energy sharing mapping method to quantify the comprehensive contribution rate of each MG to the multi-MG system, thereby achieving fair allocation of benefits. Finally, the alternating direction multiplier method is used to solve the model, effectively protecting the privacy of each MG. The simulation results demonstrate that a multi-MG system considering electricity and methanol transactions can effectively decrease carbon emissions and the total operational costs by 21.53% and 27.01% compared to only considering electricity transactions, respectively. Overall, the proposed electricity and methanol transactions strategy simultaneously reduces the overall system operation costs and carbon emissions, underscoring its advantages and significance.}, archive = {J_ESWA}, author = {Jiale Li and Bo Yang and Yiming Zhou and Hongchun Shu and Hongbiao Li and Dengke Gao and Lin Jiang}, doi = {10.1016/j.eswa.2025.129701}, journal = {Expert Systems with Applications}, month = {3}, pages = {129701}, shortjournal = {Expert Syst. Appl.}, title = {Coordinated low-carbon economic scheduling of integrated electricity-heat-gas-hydrogen-methanol multi-microgrids considering electricity-methanol transaction}, volume = {298}, year = {2026}, }
• (2026). Heterogeneous cloud resource allocation: A case study on real-time transcoding in live streaming. ESWA, 298, 129700. (www)

  @article{ , abstract = {The explosion in popularity of crowdsourced live streaming (CLS) has led to a huge increase in demand for cloud resources to support real-time video transcoding. CLS transcoding is real-time, geographically distributed and computationally intensive. Therefore, transcoding service providers need to cost-effectively utilize diverse heterogeneous cloud resources, while guaranteeing quality of service standards to ensure a good streaming experience for the viewers. To support the above, we developed a novel proactive-reactive resource allocation framework that optimizes the overall cost of supporting the CLS transcoding service using heterogeneous edge and cloud computing resources. The offline proactive policy evaluator aims to provide a good and adaptable resource usage plan in advance, matching the predicted demand with the heterogeneous resources. The reactive execution module monitors the actual demand online and controls the resource usage to compensate for deviations from the offline prediction. Our experiments show that the proposed approach leads to a cost reduction of 42 % compared to the fixed usage ratio strategy based on expert knowledge.}, archive = {J_ESWA}, author = {Yinuo Li and Jin-Kao Hao and Kwong Meng Teo and Liwei Song}, doi = {10.1016/j.eswa.2025.129700}, journal = {Expert Systems with Applications}, month = {3}, pages = {129700}, shortjournal = {Expert Syst. Appl.}, title = {Heterogeneous cloud resource allocation: A case study on real-time transcoding in live streaming}, volume = {298}, year = {2026}, }
• (2026). Adaptive multi-step path planning for multi-robot in dynamic environments based on hybrid optimization approach. ESWA, 298, 129699. (www)

  @article{ , abstract = {The multi-robot path planning problem requires algorithms with high convergence speed and accuracy, as well as the completeness of the search probability for the optimal path. The integration of metaheuristic algorithms in path planning has proven to be remarkably efficient. This paper introduces a novel hybrid metaheuristic algorithm, Beluga Whale-Crayfish Optimization (BWCOA), for enhanced global optimization in path planning applications. While the Crayfish Optimization (COA) demonstrates superior convergence speed, its inherent probabilistic path completeness remains suboptimal. To address this limitation, we present three key innovations: a dynamic probability completion mechanism, adaptive convergence acceleration factors, and balanced exploration–exploitation trade-off parameters. The proposed BWCOA synergizes Beluga Whale Optimization (BWO)’s basin-hopping capability with COA’s swarm intelligence through parallel combined exploration strategies. To prove its powerfulness, a series of comparative analyses were conducted between BWCOA and other leading algorithms across two comprehensive test function suites. The numerical experiment results underscore the significant superiority of BWCOA over its counterparts. In the context of path planning simulations, BWCOA demonstrated notable improvements over COA within the same number of function evaluations, with average enhancement rates of 6.49 %, 7.42 %, 15.09 %, 76.42 %, and 0.73 % across five evaluation metrics. Similarly, when compared to BWO on the same set of indicators, BWCOA showed average improvement rates of 22.39 %, 27.71 %, 70.53 %, 260.86 %, and 41.22 %. Furthermore, the running time of BWCOA is comparable to that of similar algorithms.}, archive = {J_ESWA}, author = {Liguo Yao and Guanghui Li and Taihua Zhang and Abdelazim G. Hussien and Yao Lu}, doi = {10.1016/j.eswa.2025.129699}, journal = {Expert Systems with Applications}, month = {3}, pages = {129699}, shortjournal = {Expert Syst. Appl.}, title = {Adaptive multi-step path planning for multi-robot in dynamic environments based on hybrid optimization approach}, volume = {298}, year = {2026}, }
• (2026). A two-stage evolutionary algorithm based on hybrid penalty strategy and its application to multi-UAV path planning. ESWA, 298, 129698. (www)

  @article{ , abstract = {Constrained optimization problems with complex and dynamic constraints pose significant challenges for evolutionary algorithms, as the constraints reshape the solution space and create conflicts between feasibility maintenance and global exploration. To address this issue, this study proposes TSC-PSODE, a two-stage evolutionary algorithm based on a hybrid penalty strategy. The algorithm employs an external penalty in the early stage to preserve population diversity and enhance exploration, while an internal penalty in the later stage accelerates convergence toward high-quality feasible solutions. In addition, a cooperative strategy combining differential evolution operators strengthens robustness and helps the population escape local optima. Experimental evaluations on the CEC2017 benchmark suite (the IEEE Congress on Evolutionary Computation 2017 benchmark) and multi-Unmanned Aerial Vehicle path planning tasks demonstrate that TSC-PSODE consistently outperforms state-of-the-art algorithms. The results confirm that the proposed method not only provides an effective mechanism for constraint handling, but also achieves a favorable balance between exploration and exploitation by maintaining diversity and accelerating convergence. In practical terms, TSC-PSODE is capable of generating safe and feasible flight paths for multiple UAVs in complex environments, highlighting its adaptability and competitiveness for real-world applications.}, archive = {J_ESWA}, author = {Eryang Guo and Yuelin Gao and Chenyang Hu}, doi = {10.1016/j.eswa.2025.129698}, journal = {Expert Systems with Applications}, month = {3}, pages = {129698}, shortjournal = {Expert Syst. Appl.}, title = {A two-stage evolutionary algorithm based on hybrid penalty strategy and its application to multi-UAV path planning}, volume = {298}, year = {2026}, }
• (2026). Critical commentary on “Reptile search algorithm (RSA): A nature-inspired meta-heuristic optimizer”. ESWA, 298, 129697. (www)

  @article{ , archive = {J_ESWA}, author = {Ngaiming Kwok}, doi = {10.1016/j.eswa.2025.129697}, journal = {Expert Systems with Applications}, month = {3}, pages = {129697}, shortjournal = {Expert Syst. Appl.}, title = {Critical commentary on “Reptile search algorithm (RSA): A nature-inspired meta-heuristic optimizer”}, volume = {298}, year = {2026}, }
• (2026). 2PADMS: Two-stage prediction and data migration strategy based on hard disk failure time. ESWA, 298, 129696. (www)

  @article{ , abstract = {With the explosive growth of global data volume, the usage of hard disk drives (HDDs) is also increasing rapidly. Consequently, the number of failed disks is continuously rising, which can affect storage service quality and even lead to data loss when failures occur.In recent years, the active fault-tolerant technology, which collects hard disks’ Self Monitoring Analysis and Reporting Technology (SMART) data-set, predicts hard disk failure by machine learning model, and repairs near-failure disks’ data to health disks in advance, has become a common research hotspot in both academia and industry. Aiming at the existing problems such as interference characteristics, inaccurate failure time prediction, competition of system resources between data migration and front service, this paper researches the two-stage prediction model and data migration strategy based on hard disk failure time, including the two-stage hard disk information feature selection method, the two-stage prediction method of hard disk failure time, and the data migration elastic system resource allocation strategy. Feature selection is performed by combining embedding methods with visualization, and the importance of the selected features is evaluated using a random forest model. Based on the feature importance, further refinement is carried out to obtain the final feature set. Before predicting the failure time of hard drives, XGBoost is first used in a voting manner to identify drives predicted to be faulty. Then, a trained Bidirectional Long Short-Term Memory network (Bidirectional LSTM) enhanced with a self-attention mechanism is employed to predict the exact failure time.Experimental results show that on the Backblaze dataset, the model achieves a mean absolute error of 1.24 when predicting failure times. The recall rate for predicting failures within 7 days reaches 98.79 %, the error rate is 0.30 %, the F1 score is 99.24 %, and the precision is 99.69 %. The elastic system resource allocation strategy for data migration improves business IOPS by 47.19 % and reduces latency by 38.68 %.}, archive = {J_ESWA}, author = {Huiyuan Qiang and Yuequan Li and Hongzhang Yang and Ping Wang and Yaofeng Tu and Shang Yang}, doi = {10.1016/j.eswa.2025.129696}, journal = {Expert Systems with Applications}, month = {3}, pages = {129696}, shortjournal = {Expert Syst. Appl.}, title = {2PADMS: Two-stage prediction and data migration strategy based on hard disk failure time}, volume = {298}, year = {2026}, }
• (2026). Federated recommender system with data valuation for E-commerce platform. ESWA, 298, 129695. (www)

  @article{ , abstract = {Federated Learning (FL) is gaining prominence in machine learning as privacy concerns grow. This paradigm allows each client (e.g., an individual online store) to train a recommendation model locally while sharing only model updates, without exposing the raw interaction logs to a central server, thereby preserving privacy in a decentralized environment. Nonetheless, despite the abundance of publicly available datasets that could substantially enrich local training, most existing FL-based recommender systems continue to rely only on private client data, leaving this potential largely unexplored. To this end, we consider a realistic scenario wherein a large shopping platform collaborates with multiple small online stores to build a global recommender system. The platform possesses global data, such as shareable user and item lists, while each store holds a portion of interaction data privately (or locally). Although integrating global data can help mitigate the limitations of sparse and biased clients’ local data, it also introduces additional challenges: simply combining all global interactions can amplify noise and irrelevant patterns, degrading personalization and increasing computational costs. To address these challenges, we propose, which selectively augments each client’s local graph with semantically aligned samples from the global dataset. employs: (i) a pre-trained graph encoder to extract global structural features, (ii) a local valid predictor to assess the client-specific relevance, and (iii) a reinforcement-learning-based probability estimator to filter and sample only the most pertinent global interactions. improves performance by up to 34.86% on recognized benchmarks in FL environments.}, archive = {J_ESWA}, author = {Jongwon Park and Minku Kang and Wooseok Sim and Soyoung Lee and Hogun Park}, doi = {10.1016/j.eswa.2025.129695}, journal = {Expert Systems with Applications}, month = {3}, pages = {129695}, shortjournal = {Expert Syst. Appl.}, title = {Federated recommender system with data valuation for E-commerce platform}, volume = {298}, year = {2026}, }
• (2026). EEG emotion recognition through a domain-adversarial multi-feature fusion network. ESWA, 298, 129694. (www)

  @article{ , abstract = {Accurate recognition of EEG signals linked to emotions is crucial for neuroscience and human-computer interaction. However, variability in EEG emotion recognition among individuals results in inconsistent feature distributions and limited generalization across subjects. To enhance the robustness of the model, we propose a deep learning approach integrating a domain adversarial migration network with an attention mechanism. Initially, a feature extractor with a hierarchical architecture (low-medium-high levels) is employed to capture multi-scale EEG features, which are then normalized for age and encoded for genderand education before being aligned with EEG features through spatio-temporal replication. Subsequently, global distribution alignment is achieved using multi-kernel maximum mean difference (MK-MMD), subdomain adversarial alignment is accomplished with a gradient inversion layer (GRL), and decision boundary clarity is enhanced through joint emotion classification loss. The generalization capability and effectiveness of the model are validated using the DEAP and DREAMER datasets, offering insights for cross-subject emotion recognition research and applications.}, archive = {J_ESWA}, author = {Weitong Sun and Yuping Su and Yumei Zhang and Xiaojun Wu}, doi = {10.1016/j.eswa.2025.129694}, journal = {Expert Systems with Applications}, month = {3}, pages = {129694}, shortjournal = {Expert Syst. Appl.}, title = {EEG emotion recognition through a domain-adversarial multi-feature fusion network}, volume = {298}, year = {2026}, }
• (2026). DFHD: Dual-granularity fusion network using historical drugs for drug recommendation. ESWA, 298, 129693. (www)

  @article{ , abstract = {Drug recommendation is a task in clinical medicine aimed at suggesting a set of safe and effective medications based on a patient’s electronic health records. Current approaches either rely on diagnoses and procedures documented in electronic health records to recommend drug combinations or focus on enhancing drug recommendation safety by considering drug-drug interactions. However, these approaches often overlook the significance of historical medication information in drug recommendation despite its strong correlation with current diagnostic and prescription recommendation. Therefore, we propose a Dual-granularity Fusion Network using Historical Drugs. Specifically, at the time-series modeling level, recurrent neural networks are used to extract time-series features from historical drug data to construct coarse-grained drug characterizations. At the molecular structure modeling level, a graph neural network is used to build a relationship map between drug molecular structures and drug substructures to capture the fine-grained interactions within drug molecules. In addition, we designed a historical drug molecule awareness module to capture historical drug information during drug molecule modeling so as to identify the drugs that really help to cure patients. To effectively integrate dual-granularity information, we design a dual-granularity fusion module to realize the synergistic learning of temporal and structural features. To ensure drug safety, we introduce the DDI loss function to adaptively adjust the loss weights based on the drug interaction risk results, taking into account the optimization goals of efficacy and safety. Our source code is available at https://github.com/AK-321/DFHD .}, archive = {J_ESWA}, author = {Kang An and Ming-Yu Lu and Yan-Kai Tian and Yi-Jia Zhang}, doi = {10.1016/j.eswa.2025.129693}, journal = {Expert Systems with Applications}, month = {3}, pages = {129693}, shortjournal = {Expert Syst. Appl.}, title = {DFHD: Dual-granularity fusion network using historical drugs for drug recommendation}, volume = {298}, year = {2026}, }
• (2026). An acoustic-electromagnetic detection system based on joint parameter estimation of atomic norm algorithm. ESWA, 298, 129692. (www)

  @article{ , abstract = {The time delay and Doppler shift parameters in radar system echo signals serve as effective tools for multi-target identification and localization in covert environments. However, the nonlinear characteristics of stationary targets are often masked by surrounding environmental factors, and traditional joint parameter estimation algorithms tend to suffer from high computational complexity and errors during demodulation. To address these challenges, this paper proposes an acoustic-electromagnetic intermodulation detection system based on a novel atomic-paradigm algorithm, which ensures localization accuracy with minimal computational complexity and zero false alarms. Specifically, the system excites the target by introducing acoustic field energy coupling, generating discernible micromotion features. The resulting acoustically modulated signal is then modeled as a two-dimensional line spectral estimation problem, capturing the target’s time delay and Doppler shift. Furthermore, the joint parameter estimation algorithm is enhanced by relaxing dyadic constraints under sufficient conditions. In our experiments, a harmonic radar physical system is constructed to simultaneously localize and measure multiple non-clustered micromotion targets. The recognition accuracy is quantitatively evaluated using a classical neural network model, achieving 86.9 % accuracy across five classified targets. The improved algorithm’s performance in joint parameter estimation is also assessed under varying signal-to-noise ratios and demodulation error rates, with a detailed time complexity analysis provided.}, archive = {J_ESWA}, author = {Sheng Wu and Yilin Cai and Yijing Zheng and Dingzhao Li and Hongjun Lai and Haixin Sun}, doi = {10.1016/j.eswa.2025.129692}, journal = {Expert Systems with Applications}, month = {3}, pages = {129692}, shortjournal = {Expert Syst. Appl.}, title = {An acoustic-electromagnetic detection system based on joint parameter estimation of atomic norm algorithm}, volume = {298}, year = {2026}, }
• (2026). A novel convex-hull-based algorithm for classification problems with imbalanced and overlapping data. ESWA, 298, 129691. (www)

  @article{ , abstract = {Classification is a fundamental task in supervised machine learning. This problem becomes challenging when dealing with imbalanced and overlapping datasets. In such cases, learning algorithms often perform well in identifying the labels of majority class data points but exhibit high error rates in predicting the minority class. This paper proposes an innovative method based on the convex-hull concept to enhance classification for imbalanced and overlapping datasets. Unlike undersampling approaches that may lead to the loss of valuable information, our method focuses on preserving the data. The process begins by clustering the data points for each class separately in such a way that no points from the opposite class fall within the convex-hull of each cluster. Then, the support vector machine (SVM) is used to separate every cluster of a given class from the data points of the opposite class. Afterward, data points inside the SVM boundaries are considered as non-overlapping, while those outside the SVM boundaries are identified as overlapping data. The XGBoost algorithm is then employed to classify the data points within the overlapping region. Extensive experiments on a variety of simulated and real-world datasets confirm the effectiveness of the proposed method in terms of various evaluation metrics, compared to existing relevant algorithms for handling imbalanced and overlapping datasets.}, archive = {J_ESWA}, author = {Farnaz Hooshmand and Sogol Peik-Mortazavi}, doi = {10.1016/j.eswa.2025.129691}, journal = {Expert Systems with Applications}, month = {3}, pages = {129691}, shortjournal = {Expert Syst. Appl.}, title = {A novel convex-hull-based algorithm for classification problems with imbalanced and overlapping data}, volume = {298}, year = {2026}, }
• (2026). Unbiased representation learning via feature decoupling network for cross-scene hyperspectral image classification. ESWA, 298, 129690. (www)

  @article{ , abstract = {For hyperspectral cross-scene classification (HSIC) tasks, the model is trained on the source domains (SDs) and applied directly to the unseen target domains (TDs). For this domain generalization (DG) challenge, a significant issue is the contradiction between the overparameterized model and the limited training domain, which results in the absorption of spurious correlations from environmental features. To alleviate this contradiction, this study proposes a domain extension generator with a feature decoupling network (FDNet). The generator initially decouples the SD into reflectance and shading components, treating them as causal and environmental features, respectively. Considering possible causal residues in environmental features, the shading components are shuffled by style to eliminate undesired correlations. Then, the reflectance and sparsified shading are reconstructed for extension, enriching the training diversity without environmental interference. In addition, to enhance the stability of class-level causal representation, a supervised aggregation strategy is designed to minimize the intra-class distance of the reflectance domain, and supervised contrastive learning is employed to enhance the class-domain semantic consistency information. Comparative analysis with advanced domain generalization and adaptation approaches on three HSI datasets validates the superiority in accuracy and Kappa coefficient metrics of the proposed method.}, archive = {J_ESWA}, author = {Hanqing Zhao and Lianlei Lin and Zongwei Zhang and Sheng Gao and Junkai Wang}, doi = {10.1016/j.eswa.2025.129690}, journal = {Expert Systems with Applications}, month = {3}, pages = {129690}, shortjournal = {Expert Syst. Appl.}, title = {Unbiased representation learning via feature decoupling network for cross-scene hyperspectral image classification}, volume = {298}, year = {2026}, }
• (2026). Automated prompt engineering pipelines: Fine-tuning LLMs for enhanced response accuracy. ESWA, 298, 129689. (www)

  @article{ , abstract = {This paper presented an Automatic Role Prompting System that seeks to improve the performance of the Large Language Model (LLM) by allowing models to assume varied roles through role-based prompting and, as a result, qualitatively improve the relevance of outputs. Our Automatic Role Prompting System’s target audience is people who do not have domain knowledge. The guiding framework (consisting of an Automated Script for discovering roles and fields layered on top of prompt engineering, and Natural Language Inference (NLI) models trained in advance), was robustly tested through the use of three datasets: our set of 1990 curated prompts, WikiQA, and the AwesomeChatGPTPrompts. We implemented a novel evaluation strategy using GPT-Eval, which scales prompts according to completeness, clarity, and relevance. We found substantially better performance than traditional rule- and template-based approaches, yielding accuracy improvements as high as 97.6 %. Overall, this work demonstrates the promise of an Automated Role Prompting System to help people engage and work more effectively and efficiently with Large Language Models (LLMs).}, archive = {J_ESWA}, author = {Samar Hendawi and Tarek Kanan and Mohammed Elbes and Ala Mughaid and Shadi Alzu’bi}, doi = {10.1016/j.eswa.2025.129689}, journal = {Expert Systems with Applications}, month = {3}, pages = {129689}, shortjournal = {Expert Syst. Appl.}, title = {Automated prompt engineering pipelines: Fine-tuning LLMs for enhanced response accuracy}, volume = {298}, year = {2026}, }
• (2026). A hybrid fusion network using convolutional vision transformers for landslide identification. ESWA, 298, 129688. (www)

  @article{ , abstract = {Convolutional neural networks (CNN) have made great strides in the segmentation of remote sensing images, but they still have certain inherent drawbacks when working with small targets and complex geological structures. These drawbacks include incomplete contextual information integration, blurry edges, and inaccurate target localisation. This study suggests using a hybrid CNN-Transformer network to improve the segmentation of landslide regions in high-resolution remote sensing images in order to overcome these difficulties. A comprehensive investigation has been conducted on the use of CNN and transformers to accomplish the task of semantic segmentation. According to experimental data, our model outperforms the state-of-the-art CNN-based, Transformer-based, and even CNN-plus-Transformer combination models for image segmentation tasks by a large margin. When it comes to landslide area segmentation, it performs exceptionally well.}, archive = {J_ESWA}, author = {S. Sreelakshmi and S.~S. Vinod Chandra}, doi = {10.1016/j.eswa.2025.129688}, journal = {Expert Systems with Applications}, month = {3}, pages = {129688}, shortjournal = {Expert Syst. Appl.}, title = {A hybrid fusion network using convolutional vision transformers for landslide identification}, volume = {298}, year = {2026}, }
• (2026). Towards intelligent online cross-selling. ESWA, 298, 129686. (www)

  @article{ , abstract = {The ubiquity of online cross-selling for fashion demands a large number of qualified outfit compositions. This paper targets practical and intelligent online cross-selling by providing a more accurate fashion compatibility model and a reliable evaluation protocol for evaluating the fashion compatibility model. Firstly, a Hierarchical Outfit Network (HON) is proposed to leverage multi-layer relations among attributes, items, and outfits. The awareness of multiple relations hidden in various outfits enables the HON to outperform all state-of-the-art methods on fill-in-the-blank (FITB) accuracy and compatibility Area Under Curve (AUC) on the Maryland and Type-aware test sets. Meanwhile, a new evaluation protocol is introduced to assess the fashion compatibility model more objectively and accurately, namely, Aesthetic 100 (A100). A100 possess three desirable characteristics: 1) Completeness . All types of standards in the fashion aesthetic system are covered through two tests, namely LAT (Liberalism Aesthetic Test) and AAT (Academicism Aesthetic Test); 2. Reliability . It is an agnostic protocol and consistent with major indicators. It provides an objective and fair assessment for model comparison. 3. Explainability . A100 assesses the model on more fine-grained dimensions, e.g. , Color, Material, and Balance demonstrating its superiority in identifying essential characteristics of fashion aesthetics. Experimental results demonstrate the progress of A100 in the aspects of Reliability and Explainability. The evaluation results on A100 also show the generalization ability of HON from both quantitative and qualitative perspectives. Finally, solutions for multiple applications in fashion retailing are proposed to show how HON can be utilized to help online cross-selling.}, archive = {J_ESWA}, author = {Kaicheng Pang and Xingxing Zou and Zowie Broach and Waikeung Wong}, doi = {10.1016/j.eswa.2025.129686}, journal = {Expert Systems with Applications}, month = {3}, pages = {129686}, shortjournal = {Expert Syst. Appl.}, title = {Towards intelligent online cross-selling}, volume = {298}, year = {2026}, }
• (2026). GCMF-net: Global-aware cross-attention and mask-guided fusion for multispectral object detection. ESWA, 298, 129679. (www)

  @article{ , abstract = {Multispectral Object Detection has shown significant advantages under various lighting and weather conditions, with efficient fusion of RGB and thermal information playing a key role. Previous studies have demonstrated the effectiveness of feature fusion based on convolutional neural networks, but limited local feature interactions hinder the capture of global complementary information. To address these limitations, some methods adopt complex fusion strategies to enhance complementary feature extraction, yet fail to mitigate feature redundancy, which negatively impacts detection performance. In this paper, we propose a novel Global-aware Cross-attention and Mask-guided Fusion (GCMF) module for Multispectral Object Detection, following a fusion-refinement paradigm. In the fusion stage, we first introduce Efficient Channel Attention with Weighted Max-Pooling (ECA-WM) to focus on key information within each modality and achieve implicit alignment before fusion. Subsequently, the Global-aware Cross-Attention Transformer (GCAT) effectively captures complementary cross-modal information and models global features. In the refinement stage, the Mask-guided Refinement Strategy (MRS) generates segmentation masks to distinguish target features from the background. These masks are applied before and after cross-modal interaction to highlight target-relevant information while suppressing irrelevant features, resulting in highly discriminative fused representations. Extensive experimental comparisons demonstrate that the proposed GCMF fusion strategy achieves state-of-the-art performance on the publicly available FLIR, LLVIP and DVTOD datasets, with absolute improvements of 2.8 %, 4.2 % and 4.0 % over the previous methods, respectively. Moreover, the proposed strategy is general and effective, making it adaptable to various detection frameworks.}, archive = {J_ESWA}, author = {Zhong Qu and Jin Yang and Shufang Xia}, doi = {10.1016/j.eswa.2025.129679}, journal = {Expert Systems with Applications}, month = {3}, pages = {129679}, shortjournal = {Expert Syst. Appl.}, title = {GCMF-net: Global-aware cross-attention and mask-guided fusion for multispectral object detection}, volume = {298}, year = {2026}, }
• (2026). Oriented bounding box detection algorithm for dense scenarios of robotic arm operation. ESWA, 298, 129678. (www)

  @article{ , abstract = {In the context of the national promotion of intelligent manufacturing and the ’Industry4.0’ strategy, the demand for intelligent robotic arms in industrial production has steadily increased. However, challenges such as occlusion, significant object scale variations, and strict real-time requirements have made target detection in densely packed environments more challenging. This study, based on the YOLOv11 algorithm, proposes an efficient oriented bounding box detection method aimed at improving the model’s performance in feature extraction, computational efficiency, and network lightweighting to tackle target detection challenges in dense industrial settings. A Dynamic-Cross-Stage-Dual-Conv module was designed to enhance the Bottleneck section, employing a parallel dual-branch structure for local feature extraction and global context fusion. Simultaneously, a CIoU loss function with geometric perception was introduced to improve object localization accuracy and strengthen the network’s ability to handle densely stacked objects. Next, a Modulated-Deform-Conv deformable convolution module was integrated into the Backbone structure, dynamically adjusting the convolution kernel sampling positions, enabling the network to learn deformation features in dense scenes, improving adaptability to shape and scale variations while reducing computational load. Additionally, a C3K2_FasterBlock lightweight structure, utilizing partial convolutions and sparse connections, was proposed to minimize redundant calculations and optimize feature interactions. On a custom-built dense object dataset, the model achieved a 2.9 % increase in mAP@0.5 and reduced computational cost by 14 %. Finally, the improved model was deployed on the Jetson Orin Nano development board, demonstrating its practical value in robotic arm recognition and grasping tasks in dense industrial environments, offering a new paradigm for future applications.}, archive = {J_ESWA}, author = {Jinshun Dong and Lixia Deng and Dapeng Wan and Chenxu Liu and Jianqin Yin and Meiqi Guo and Hongyu Zhang and Shoujun Lin and Haiying Liu and Lida Liu}, doi = {10.1016/j.eswa.2025.129678}, journal = {Expert Systems with Applications}, month = {3}, pages = {129678}, shortjournal = {Expert Syst. Appl.}, title = {Oriented bounding box detection algorithm for dense scenarios of robotic arm operation}, volume = {298}, year = {2026}, }
• (2026). CrossModalNet: A dual-modal object detection network based on cross-modal fusion and channel interaction. ESWA, 298, 129677. (www)

  @article{ , abstract = {Object detection in complex environments, particularly under conditions such as low illumination or adverse weather, remains a significant challenge. Multispectral detection techniques that integrate visible and infrared imagery offer a promising solution by leveraging complementary modality information. However, substantial discrepancies between these modalities hinder traditional fusion strategies, which often fail to adaptively align and integrate features, resulting in information loss and diminished detection performance. To overcome this limitation, we propose CrossModalNet, a novel cross-modal fusion and channel interaction framework. CrossModalNet comprises two key modules: Convolutional Attention Interaction Module (CAIM) and Bidirectional Cross-Modal Attention Module (BiCAM). CAIM enables effective cross-modal integration across varying semantic levels by employing convolutional attention mechanisms combined with pixel-wise channel guidance. In parallel, BiCAM enhances inter-modal feature complementarity by modeling channel-level interactions through bidirectional attention pathways. Extensive experiments conducted on four public multispectral datasets, FLIR, LLVIP, M3FD, and VEDAI, demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches across multiple performance metrics. Moreover, with an inference speed of 12.6 FPS on embedded platforms, the proposed model is suitable for real-time deployment.}, archive = {J_ESWA}, author = {Hanyun Li and Linsong Xiao and Lihua Cao and Di Wu and Yangfan Liu and Yi Li and Yunfeng Zhang and Haiyang Bao}, doi = {10.1016/j.eswa.2025.129677}, journal = {Expert Systems with Applications}, month = {3}, pages = {129677}, shortjournal = {Expert Syst. Appl.}, title = {CrossModalNet: A dual-modal object detection network based on cross-modal fusion and channel interaction}, volume = {298}, year = {2026}, }
• (2026). IPMMG: Information propagation with multi-granularity morphology-guided for nuclear segmentation and classification. ESWA, 298, 129676. (www)

  @article{ , abstract = {Nuclear segmentation and classification play a crucial role in pathological image analysis. However, it is frequently challenged by blurred nuclear boundaries and complex structures in digital pathology slides, due to factors such as staining techniques and imaging methods, posing a significant challenge for accurate segmentation and classification. To this end, we propose a novel and efficient approach for nuclear identification, termed Information Propagation with Multi-Granularity Morphology-Guided Network (IPMMG). Specifically, IPMMG progressively captures edge morphology information from different network layers while simultaneously incorporating structural morphology features at multiple granularities. By explicitly propagating features related to both the edge and the structure, our approach constrains semantic features to focus on contours of the region of interest in the nuclear segmentation task, thus mitigating the challenge of blurred morphology. Experiments on public datasets demonstrate that IPMMG achieves state-of-the-art (SOTA) performance in segmentation, as measured by Dice and IoU scores, while also attaining competitive results in classification with DQ, SQ, and PQ metrics. In particular, our proposal IPMMG excels in handling nuclei with blurred edges and complex structures.}, archive = {J_ESWA}, author = {Dawei Fan and Jun Li and Chengfei Cai and Lihui Lin and Riqing Chen and Yanping Chen and Lifang Wei}, doi = {10.1016/j.eswa.2025.129676}, journal = {Expert Systems with Applications}, month = {3}, pages = {129676}, shortjournal = {Expert Syst. Appl.}, title = {IPMMG: Information propagation with multi-granularity morphology-guided for nuclear segmentation and classification}, volume = {298}, year = {2026}, }
• (2026). MECA: Modular editing via customized expert networks and adaptors in large language models. ESWA, 298, 129675. (www)

  @article{ , abstract = {Updating language models with new information through targeted edits without resorting to expensive full model retraining remains a critical challenge, particularly when aiming to preserve pre-existing capabilities. In this work, we introduce M odular E diting via C ustomized expert networks and A daptors) (MECA), a unified framework that selectively integrates new knowledge into language models. MECA employs a module-level deferral router to evaluate whether incoming queries fall within the scope of existing edit requests. Queries are then dynamically routed to either customized editing experts or key-value adaptors. This modular strategy ensures that updates are localized, thereby mitigating risks of unintended alterations on unrelated outputs. We validate our approach on sequential editing tasks using Llama2-7B, Llama2-13B and Falcon 11B, benchmarked across two diverse datasets ZsRE and Hallucination. Experimental results show that MECA consistently outperforms several state-of-the-art knowledge editing techniques, achieving improved integration of new information while preserving the model’s original performance. Our analysis further demonstrates that the deferral routing mechanism for selecting modules effectively balances editing precision with overall model stability.}, archive = {J_ESWA}, author = {Roseline Nyange and Shanbao Qiao and Seung Hoon Na}, doi = {10.1016/j.eswa.2025.129675}, journal = {Expert Systems with Applications}, month = {3}, pages = {129675}, shortjournal = {Expert Syst. Appl.}, title = {MECA: Modular editing via customized expert networks and adaptors in large language models}, volume = {298}, year = {2026}, }
• (2026). An explainable hybrid bio-inspired feature selection framework using RSVP-evoked p300 EEG signals for identity authentication. ESWA, 298, 129674. (www)

  @article{ , abstract = {EEG-based personal identification confronts critical hurdles, including high dimensionality, noise, and real-time variability. While RSVP and P300 paradigms provide cognitive-response-driven security, feature extraction challenges prevent practical deployment. Although deep learning has addressed unstructured EEG data, pinpointing optimal RSVP and P300-specific features remains an unresolved issue. To overcome these limitations, we introduced a hybrid GWO-MSE-XAI framework integrating Grey Wolf Optimization (GWO), Multiscale Entropy (MSE), and SHAP-based Explainable AI (XAI) to select the most relevant features from RSVP-evoked P300 EEG signals. The framework prioritizes discriminative feature selection, improves class separability, and incorporates a hybrid cross-entropy loss function fused with Fisher’s score-based feature selection. Benchmark-driven optimization refines EEG-specific feature subsets, while evaluation using classifiers (Random Forest, LightGBM, CatBoost, XGBoost) demonstrates substantial dimensionality reduction, faster convergence, and superior performance (98.89% accuracy). Experimental results confirm robustness, scalability, and enhanced interpretability, positioning the framework as a viable solution for EEG-based identity authentication in real-world RSVP and P300 applications.}, archive = {J_ESWA}, author = {S Abinayaa and S.S. Sridhar}, doi = {10.1016/j.eswa.2025.129674}, journal = {Expert Systems with Applications}, month = {3}, pages = {129674}, shortjournal = {Expert Syst. Appl.}, title = {An explainable hybrid bio-inspired feature selection framework using RSVP-evoked p300 EEG signals for identity authentication}, volume = {298}, year = {2026}, }
• (2026). Interlink reconfiguration against cascading failures on cyber-physical power systems based on an improved memetic algorithm. ESWA, 298, 129673. (www)

  @article{ , abstract = {Cyber-physical power systems (CPPSs) are the deep integration of advanced information and other technologies applied to the grid to achieve fundamental changes in the power industry. Thus, it is very significant to enhance the robustness of the power communication system in order to ensure security. For instance, when an interdependent CPPSs is under attack, the failure can diffuse along interconnect topology to the whole system, even causing a system crash. Recall that most of the existing models are one-to-one coupling structures. Drawing close to reality, we develop a modified memetic algorithm to reconstruct CPPSs with multiple-to-multiple interlinks, called MA-Multiple, which is dedicated to improving the robustness of CPPSs. To improve the accuracy of the optimal solution of the MA-Multiple, we devise a crossover operator (CO) based on the null model to increase population diversity. Further improving the suitability of algorithms on CPPSs, we propose a new Kirchhoff centrality (KIC) based on the approximation of the inverse matrix to measure network connectivity and design a local search operator (LSO). In the experiment, we comprehensively compare the MA-Multiple with four closely relevant methods in different scenarios. The results imply that MA-Multiple performs better in enhancing network robustness and resisting attacks.}, archive = {J_ESWA}, author = {Wei Lin and Li Xu and Yuexin Zhang and Jie Li}, doi = {10.1016/j.eswa.2025.129673}, journal = {Expert Systems with Applications}, month = {3}, pages = {129673}, shortjournal = {Expert Syst. Appl.}, title = {Interlink reconfiguration against cascading failures on cyber-physical power systems based on an improved memetic algorithm}, volume = {298}, year = {2026}, }
• (2026). Multi-objective portfolio optimization for stock return prediction using machine learning. ESWA, 298, 129672. (www)

  @article{ , abstract = {This paper presents a novel approach that integrates stock return prediction with the mean–variance (MV) model to enhance the performance of the original model. Firstly, stock returns are predicted using machine learning algorithms, including Robust Linear Regression (OLS-H), Random Forest (RF), and Long Short-Term Memory Networks (LSTM), to select a pre-screened stock pool composed of stocks with high predicted returns. Secondly, a linear weighting method combines the predictions above with the MV model, constructing the Mean-Variance-Forecast Error (MVF) model and determining the investment proportions for the pre-selected stocks. Finally, empirical research is conducted using the components of the CSI 300 Index as sample data. The results indicate that the RF + MVF model outperforms other models and the CSI 300 Index in return and risk metrics. At the same time, a sensitivity analysis of relevant parameters further confirms that considering return uncertainty is beneficial for improving the out-of-sample performance of the MV model.}, archive = {J_ESWA}, author = {Meiyu Huang and Shili Dang and Miraj Ahmed Bhuiyan}, doi = {10.1016/j.eswa.2025.129672}, journal = {Expert Systems with Applications}, month = {3}, pages = {129672}, shortjournal = {Expert Syst. Appl.}, title = {Multi-objective portfolio optimization for stock return prediction using machine learning}, volume = {298}, year = {2026}, }
• (2026). EKT-ML: An efficient knowledge tracing model with multi-task learning. ESWA, 298, 129671. (www)

  @article{ , abstract = {Knowledge Tracing (KT) aims to trace a student’s mastery of knowledge, known as knowledge states, and has become a popular research area, with Self-Attention (SA)-based KT models achieving the state-of-the-art performance. However, existing SA-based KT models seem to still have issues that need further investigation. Firstly, there commonly exists incorrect question-knowledge concept (Q-KC) mapping, yet most models fail to address this issue. Secondly, existing SA-based KT models suffer from high time complexity due to their extensive use of the SA mechanism. Finally, from real-world datasets, we observe that there exists a repeated attempts pattern which is often overlooked by existing KT models. Motivated by the above observations, we propose a novel Efficient Knowledge Tracing Model with Multi-task Learning (EKT-ML), an SA-based model with three crucial features. Firstly, we formulate the KT as a Multi-task Learning, with Q-task and KC-task as two tasks; by training them simultaneously and treating Q-KC mapping as shared information, the proposed EKT-ML tends to mitigate the impact of incorrect Q-KC mapping. Moreover, we propose an SVD-MLP component to replace the initial SA layers commonly used in existing SA-based KT models, thereby reducing the time complexity of EKT-ML. Finally, experimental results show that EKT-ML improves performance by up to 8.84 % across four metrics on three widely used datasets. Furthermore, it demonstrates that the EKT-ML has reduced time complexity compared with the benchmark baseline.}, archive = {J_ESWA}, author = {Wei Liu and Bo Yang and Haotian Su and Yaowei Wang and Qing Li}, doi = {10.1016/j.eswa.2025.129671}, journal = {Expert Systems with Applications}, month = {3}, pages = {129671}, shortjournal = {Expert Syst. Appl.}, title = {EKT-ML: An efficient knowledge tracing model with multi-task learning}, volume = {298}, year = {2026}, }
• (2026). Attention dilated deep learning network-based distributed data mining with energy efficient heuristic-aided routing model in WSN. ESWA, 298, 129670. (www)

  @article{ , abstract = {Wireless Sensor Networks (WSNs) are comprised of autonomous nodes, where the physical devices and other equipment are embedded for monitoring the environment. Yet, it consumes larger energy consumption while it fails to provide the optimal routing path. The data aggregation in WSN is accomplished by several classification approaches, but the delay and utilization of the resource are high. So, it is necessary to solve the challenges by implementing the developed model for the energy efficient routing in WSN. A distributed data mining scheme is developed based on a deep learning approach to provide greater energy efficiency and optimal load balancing in WSN. The distributed data mining is performed to distribute the appropriate data to the sensor nodes. This data distribution highly reduces the overhead at the fusion center. The data are categorized via Fused Deep Structural Network (FDSNet), where the Attention-based Dilated Deep Temporal Context Networks (DTCN) are fused with a Deep Shallow Network. The necessary data are distributed to the sensor nodes. The optimal routing is carried out based on optimal path selection using Stability Bound-based Energy Valley Optimizer (SBEVO). The optimal routing is done via the multi-objective functions such as “energy consumption, PDR, delay, and shortest path”. The energy-efficient data transmission is accomplished with the help of developed distributed data mining with a routing scheme. The developed energy-efficient optimal routing in WSN is compared to conventional routing approaches with several performance metrics to show energy efficiency. While comparing with existing methods, the developed model attains the value of 0.95% with regards of accuracy, precision, and F1-score. These findings of the developed model show accurate performance with an efficient routing mechanism in WSN.}, archive = {J_ESWA}, author = {Kiran Goud Palakuri and Manjeet Singh}, doi = {10.1016/j.eswa.2025.129670}, journal = {Expert Systems with Applications}, month = {3}, pages = {129670}, shortjournal = {Expert Syst. Appl.}, title = {Attention dilated deep learning network-based distributed data mining with energy efficient heuristic-aided routing model in WSN}, volume = {298}, year = {2026}, }
• (2026). Curriculum learning for a hybrid approach for aspect-based sentiment analysis. ESWA, 298, 129669. (www)

  @article{ , abstract = {In the past years, the amount of unstructured online review data has grown exponentially. Many people express their opinions about different aspects of goods and services on the Web. Aspect-Based Sentiment Analysis (ABSA) automatically extracts the sentiments with respect to aspects given in a sentence. We improve the training procedure of the state-of-the-art Hybrid Approach for Aspect-Based Sentiment Analysis with deep contextual word embeddings and hierarchical attention (HAABSA++). In this method, a domain sentiment ontology is used as a main classifier, and if it is not conclusive, a neural network is employed as a back-up. We extend the training of the neural network by incrementally adding more difficult instances, also known as curriculum learning. Restaurant reviews obtained from the SemEval-2015 and SemEval-2016 datasets are used to evaluate the effect of implementing curriculum learning. Using baby steps curriculum learning and a specific curriculum strategy, the accuracy of HAABSA++ is improved from 86.3 % to 87.5 %.}, archive = {J_ESWA}, author = {Nana Lange and Flavius Frasincar and Maria Mihaela Truşcǎ}, doi = {10.1016/j.eswa.2025.129669}, journal = {Expert Systems with Applications}, month = {3}, pages = {129669}, shortjournal = {Expert Syst. Appl.}, title = {Curriculum learning for a hybrid approach for aspect-based sentiment analysis}, volume = {298}, year = {2026}, }
• (2026). A new edge detection method for noisy image based on discrete fractional wavelet transform and improved canny algorithm. ESWA, 298, 129668. (www)

  @article{ , abstract = {Edge detection is a key technique for extracting object structures and region boundaries in images, serving as an important foundation for visual tasks such as image segmentation and object recognition. However, during real-world image acquisition, various types of noise are inevitably introduced into the images. Traditional edge detection methods suffer significant performance degradation in noisy environments, often resulting in false edges or missing true edges. To address this issue, this paper proposes a novel edge detection method for noisy images. The method begins by adaptively selecting an optimal fractional order p , based on the distribution characteristics of the image’s subband modulus coefficients. This order is then used to perform a p -order discrete fractional wavelet transform (DFRWT) on the noisy image. Then, within the DFRWT domain, an enhanced Canny algorithm is applied to detect edges. This algorithm improves upon the standard method by replacing the traditional gradient operator with a more robust fractional-order Sobel operator to compute the gradient magnitude. This detection process is performed on both the low- and high-frequency subbands to capture features at different scales. Finally, the edge images from the low- and high-frequency components are reconstructed to obtain the final edge detection result. Experimental results demonstrate that, compared to four representative edge detection algorithms, the proposed method exhibits superior noise robustness and edge preservation capability in noisy environments.}, archive = {J_ESWA}, author = {Xiaozhong Yang and Chunmeng Li}, doi = {10.1016/j.eswa.2025.129668}, journal = {Expert Systems with Applications}, month = {3}, pages = {129668}, shortjournal = {Expert Syst. Appl.}, title = {A new edge detection method for noisy image based on discrete fractional wavelet transform and improved canny algorithm}, volume = {298}, year = {2026}, }
• (2026). Physics-constrained EBSD image inpainting via adversarial graph learning: Bridging crystallographic rules and multimodal deep learning. ESWA, 298, 129667. (www)

  @article{ , abstract = {Electron Backscatter Diffraction (EBSD) is a crucial characterisation method in materials engineering. The reliability of EBSD data is essential in the aerospace, nuclear, and automotive industries, as material performance greatly affects operational safety. While industrial practice makes perfect EBSD data difficult, with sample preparation errors, beam drift, and instrumental noise corrupting up to one-third of datasets. Automated crystallographic fidelity restoration solutions are needed because corrupted data force engineers to abandon valuable experiments or manually restore datasets at risk of errors. Current image inpainting techniques fail to maintain crystallographic constraints, resulting in restorations that violate the basic rules for crystalline materials. A novel physics-constrained framework is proposed to fill this gap. It integrates adversarial learning with graph neural networks (GNNs) for crystallographically consistent EBSD image inpainting. The proposed GTRG method consists of three elements: i ) a generative adversarial network (GAN) for reconstructing grain boundaries; i i ) a crystallography-guided graph transformer (T) that converts pixel data into orientation-boundary graphs; and i i i ) a regression graph convolutional network (RGCN) that links grain, orientation and boundaries to predict missing crystal orientations. The framework mandates a single orientation per grain and preserves grain boundary structure through structured graph representations. A strategy for creating automated EBSD datasets that incorporates realistic corruption patterns supports effective model training and evaluation. Experimental validation shows better performance than current methods, with a 3.5 % improvement in SSIM (0.950 vs. 0.918) and a 63.0 % reduction in FID (16.55 vs. 44.70) compared to AOT-GAN. The study on aerospace niobium alloys further validates practical utility, showing statistically consistent grain orientation and size distributions (Kolmogorov-Smirnov D = 0.02 , p > 0.98 ). This work introduces two key advancements: 1) the first integration of graph neural networks with adversarial learning for topology-aware image inpainting, and 2) a physics-informed framework bridging computer vision and materials science, enabling effective restoration of corrupted EBSD data for subsequent engineering applications.}, archive = {J_ESWA}, author = {Baiyang Zheng and Jiongran Wen and Yat-Sze Choy and Chengwei Fei}, doi = {10.1016/j.eswa.2025.129667}, journal = {Expert Systems with Applications}, month = {3}, pages = {129667}, shortjournal = {Expert Syst. Appl.}, title = {Physics-constrained EBSD image inpainting via adversarial graph learning: Bridging crystallographic rules and multimodal deep learning}, volume = {298}, year = {2026}, }
• (2026). Three-way large-scale group decision-making under incomplete multi-scale information systems: A perspective of quantum social networks. ESWA, 298, 129666. (www)

  @article{ , abstract = {The gathering and sharing of information lay the groundwork for decision-making, while large-scale group decision-making (LSGDM) strategies address biases, promoting a more comprehensive evaluation of alternatives. Regarding information representation, incomplete multi-scale information systems (MSISs), as an application of granular computing, combine inputs from decision-makers (DMs) and tackle data gaps through multi-level analysis to foster LSGDM. Furthermore, given the interference effect among DMs, quantum social networks (SNs) and three-way decisions (TWD) are vital for effective decision-making. Quantum SNs provide a framework for modeling complex trust relationships among DMs, while TWD offers a structured approach to manage uncertainty. Therefore, this paper seeks to investigate quantum SN-guided three-way LSGDM under incomplete MSISs. First, MSISs are designed to gather information across spatial dimensions. Second, trust propagation paths within SNs are aggregated using quantum theory. Following community clustering through the Leiden algorithm, each community is further divided into core and fringe regions by three-way clustering (TWC), where core alternatives reflect the central members and fringe alternatives represent uncertain members. Third, to achieve intra-group consensus, the weights of DMs in fringe regions and those with low consensus levels are adjusted, while for inter-group consensus, the weight and decision information of community representatives with low consensus levels are modified. Fourth, alternatives are classified using the TWD method, which is grounded in the Dempster-Shafer theory and incorporates the enhanced belief Jensen-Sharma-Mittal ( E B J S M ) divergence. Finally, air quality datasets are used to validate the practicality of this method through sensitivity analysis, simulation analysis, comparative analysis, and statistical analysis.}, archive = {J_ESWA}, author = {Rui Li and Chao Zhang and Hamido Fujita and Wentao Li and Witold Pedrycz and Oscar Castillo}, doi = {10.1016/j.eswa.2025.129666}, journal = {Expert Systems with Applications}, month = {3}, pages = {129666}, shortjournal = {Expert Syst. Appl.}, title = {Three-way large-scale group decision-making under incomplete multi-scale information systems: A perspective of quantum social networks}, volume = {298}, year = {2026}, }
• (2026). GDCR: Geometry-enhanced directional consistency representation for point cloud analysis. ESWA, 298, 129665. (www)

  @article{ , abstract = {Point clouds provide discrete representations of 3D scenes. The relative positions and directions between points collectively describe the objects. Variations in sampling angles, distances, or noises can introduce perturbations, disrupting these spatial and directional relationships. These pose significant challenges for achieving robust feature representations. However, research on the robust representation of point clouds is limited. Although advanced models achieve impressive performance, they exhibit poor robustness to perturbations. To address this issue, we propose Geometry-enhanced Directional Consistency Representation (GDCR), a novel method designed to enhance robustness. In GDCR, we introduce Statistic-based Geometric Reasoning (SGR) to achieve precise spatial geometric estimation for discrete point sets, explicitly enriching spatial geometric information. Furthermore, GDCR vectorizes features embedded with SGR information and applies feature rotation and relative direction refinement in the expanded feature space for robust directional representation. GDCR improves the flexibility and directional expressiveness of point cloud features, significantly improving robustness against perturbations. Extensive experiments demonstrate that GDCR exhibits outstanding robustness while surpassing or matching the performance of state-of-the-art methods.}, archive = {J_ESWA}, author = {Ziming Wang and Boxiang Zhang and Ming Ma and Yue Wang and Taoli Du and Ying Wang and Wenhui Li}, doi = {10.1016/j.eswa.2025.129665}, journal = {Expert Systems with Applications}, month = {3}, pages = {129665}, shortjournal = {Expert Syst. Appl.}, title = {GDCR: Geometry-enhanced directional consistency representation for point cloud analysis}, volume = {298}, year = {2026}, }
• (2026). APMoE-net: Fourier amplitude-phase joint enhancement and MoE compensation for low-light image enhancement. ESWA, 298, 129664. (www)

  @article{ , abstract = {Low-Light Image Enhancement (LLIE) plays a crucial role in computer vision applications. Beyond spatial-based approaches, recent works have explored the Fourier domain. To better preserve structural details in extremely dark scenes, infrared modality has also been introduced as a robust prior for capturing scene geometry. However, existing methods suffer from limited enhancement performance due to the independent modeling of Fourier amplitude and phase, the limitations of cross-modal guidance, and the information loss in sequential feature extraction. To address these challenges, we propose APMoE-Net, a dual-stage Fourier network framework with amplitude-phase joint enhancement and spatial Mixture of Experts (MoE) compensation. Stage one performs coarse enhancement by leveraging infrared images to jointly optimize Fourier amplitude and phase, enabling mutual guidance learning between them. Subsequently, a Modality Refinement Module leverages edge information to refine infrared inputs, producing a refined modality map as a more accurate cross-modal prior for subsequent processing. The second stage employs a dual-branch design for texture refinement. Our key innovation lies in the MoE Compensation Module integrated within the Multi-scale Convolution Branch. This module employs a dynamic routing network to selectively activate specialized experts, enabling the recovery of fine-grained textures that are lost during sequential processing. Meanwhile, the Fourier Branch integrates the refined modality map to improve overall detail and contrast. Comprehensive experiments demonstrate that APMoE-Net surpasses state-of-the-art (SOTA) methods in both qualitative and quantitative evaluations. Notably, APMoE-Net achieves outstanding performance with a lightweight design, offering an efficient LLIE solution.}, archive = {J_ESWA}, author = {Mengen Cai and Tongshun Zhang and Pingping Liu and Qiuzhan Zhou}, doi = {10.1016/j.eswa.2025.129664}, journal = {Expert Systems with Applications}, month = {3}, pages = {129664}, shortjournal = {Expert Syst. Appl.}, title = {APMoE-net: Fourier amplitude-phase joint enhancement and MoE compensation for low-light image enhancement}, volume = {298}, year = {2026}, }
• (2026). Enhanced edge detection of harmful algal blooms using diffusion probability models and sobel-convolutional attention mechanisms. ESWA, 298, 129663. (www)

  @article{ , abstract = {Accurate detection and identification of harmful algal bloom (HAB) images are crucial for developing effective early warning systems for HABs. However, existing edge detection models, primarily designed for natural scenes, struggle with HAB-specific challenges such as blurred cell contours and interference from impurity bubbles. To address these issues, we propose a novel edge detection approach tailored for marine HABs, integrating a diffusion probability model with Sobel convolutional inter-layer attention mechanisms. Firstly, we develop an image enhancement algorithm specifically for HABs images, significantly improving real-time dynamic sampling data by enhancing contrast, edges, and texture features. Next, we introduce the SIAnet network, which utilizes inter-layer attention and convolutional operations to generate comprehensive global information. This network enhances feature correlation by aggregating shared features across multiple layers and modeling both long-range and short-range dependencies, effectively suppressing noise and background interference. This facilitates precise extraction of algae boundaries and morphological characteristics. Additionally, an improved Sobel operator is employed to generate supplementary edge features, accelerating the training process. Experimental results demonstrate that the proposed method achieves robust performance on the HABs dataset, with an Optimal Dataset Scale (ODS) of 0.645, an Optimal Image Scale (OIS) of 0.702, and an Average Precision (AP) of 0.813. Compared to existing methodologies, our approach demonstrates strong generalization on BSDS and BIPED datasets, significantly enhancing performance and mitigating typical CNN issues of edge thickening and fragmentation. It offers essential technical support for efficient HAB early warning system development.}, archive = {J_ESWA}, author = {Gengkun Wu and Yining Fan and Xin Tian and Chao Cui and Jiazheng Han}, doi = {10.1016/j.eswa.2025.129663}, journal = {Expert Systems with Applications}, month = {3}, pages = {129663}, shortjournal = {Expert Syst. Appl.}, title = {Enhanced edge detection of harmful algal blooms using diffusion probability models and sobel-convolutional attention mechanisms}, volume = {298}, year = {2026}, }
• (2026). Predictive modeling of biogeographical ancestry using a novel SNP panel and supervised learning approaches. ESWA, 298, 129662. (www)

  @article{ , abstract = {Inferring an individual’s BioGeographical Ancestry (BGA) through DNA analysis is a valuable tool in various fields such as forensic science, especially when traditional methods fail to identify suspects or victims. Advances in Next-Generation Sequencing (NGS) have revolutionized genomic data acquisition, enabling the development of comprehensive Single Nucleotide Polymorphism (SNP) panels for ancestry inference. This study assessed the effectiveness of a novel panel containing 3234 SNPs at both inter-continental and a more detailed BGA level, using various supervised Machine Learning (ML) models, including Categorical Naive Bayes, Penalized Multinomial Logistic Regression, Linear Support Vector Machines, Random Forest, and tree-based Gradient Boosting. A nested cross-validation approach was employed for model tuning and evaluation, with balanced accuracy as the main performance metric to address class imbalance. At the inter-continental level, all ML models demonstrated high balanced accuracy, confirming their reliability for BGA inference. However, performance declined at the more detailed continental level, likely due to a combination of factors including increased class imbalance, reduced sample sizes for certain populations, and the inherent complexity of distinguishing genetically and geographically proximate groups. Nonetheless, promising results were observed for South Asians, Northeast Asians, Europeans, and West Africans classes. In contrast, performance was notably lower for underrepresented classes such as Inner Asians. Misclassification patterns at both levels appeared to reflect known geographical and historical relationships, although further analysis revealed that these were often concentrated in underrepresented or genetically complex groups. These findings highlighted the potential of this SNP panel and ML approaches as valuable tools for forensic investigations.}, archive = {J_ESWA}, author = {Cosimo Grazzini and Giorgia Spera and Stefania Morelli and Daniele Castellana and Giulia Cosenza and Michela Baccini and Giulia Cereda and Elena Pilli}, doi = {10.1016/j.eswa.2025.129662}, journal = {Expert Systems with Applications}, month = {3}, pages = {129662}, shortjournal = {Expert Syst. Appl.}, title = {Predictive modeling of biogeographical ancestry using a novel SNP panel and supervised learning approaches}, volume = {298}, year = {2026}, }
• (2026). RL-CoSeg: A reinforcement learning-based collaborative localization and segmentation framework for medical image. ESWA, 298, 129661. (www)

  @article{ , abstract = {Segmenting small regions of interest (ROIs) from abdominal CT images presents significant challenges, particularly due to class imbalance and variations in the sizes of foreground objects. A commonly adopted solution is the two-stage segmentation. However, this approach has two key limitations: i) Difficulty in balancing localization accuracy and target preservation. To reduce information loss in the first stage, existing methods typically enlarge the predicted bounding boxes, which improves coverage but compromises localization precision. ii) Independent optimization of the two stages, which lacks a collaborative mechanism. This fragmented pipeline limits the flow of information between stages, thereby constraining performance improvements. To address these limitations, we propose a reinforcement learning-based collaborative localization and segmentation (RL-CoSeg) framework, which comprises three sub-networks: localization network (LN), segmentation network, and localization-segmentation collaboration network (LSCN). The LN integrates prior knowledge and incorporates a dynamic reward mechanism to enhance the accuracy and efficiency of target detection through reinforcement learning (RL) strategies. The LSCN further introduces segmentation predictions as a reward signal, which, together with the localization reward, jointly drives policy learning. In addition, a heuristic exploration strategy is employed to avoid local optima and improve training stability. This design strengthens the information interaction and collaborative performance between the two tasks. Experimental results demonstrate that the proposed method achieves superior collaborative performance in small-target medical image segmentation.}, archive = {J_ESWA}, author = {Feilong Xu and Feiyang Yang and Xiaoli Zhang and Zhaojun Liu}, doi = {10.1016/j.eswa.2025.129661}, journal = {Expert Systems with Applications}, month = {3}, pages = {129661}, shortjournal = {Expert Syst. Appl.}, title = {RL-CoSeg: A reinforcement learning-based collaborative localization and segmentation framework for medical image}, volume = {298}, year = {2026}, }
• (2026). SH-ETRs:Learning soft-hard rules with entity type constraints for document-level relation extraction. ESWA, 298, 129660. (www)

  @article{ , abstract = {Document-level relation extraction (DocRE) aims to extract relations between entity pairs across the entirety of a document. Current methods have begun to adopt logical rules to enhance the performance of DocRE models. However, the pipeline’s rule learning framework will suffer from the issue of error propagation, and the end-to-end method may lead to mistakes in rule reasoning. Additionly, they ignore entity type information when learning the rules. To address these issues, we propose a novel framework named Soft-Hard Rules with Entity Type Constraints (SH-ETRs) for improving the rules’ expressiveness and quality. Specifically, we first propose a Hard Entity Type Rules Module (H-ETRs) to learn entity type information and provide hard rule constraints. Then, we propose a Soft Entity Type Rule Reasoning Module (S-ETRs), which parameterizes the rule inference process and reduces error propagation during the process. Furthermore, by applying a rule consistency loss function to S-ETRs, we achieve the learning of soft rules under hard rule constraints, thereby aiming to prevent the learning of inaccurate rules during the training process. The experimental results demonstrate that our method outperforms existing rule learning frameworks, achieving state-of-the-art performance with an F1 score of 74.39 and an IgnF1 score of 67.43 across three public datasets and two baseline models.}, archive = {J_ESWA}, author = {Haisong Chen and Nisuo Du and Qing He and Yuji Wang}, doi = {10.1016/j.eswa.2025.129660}, journal = {Expert Systems with Applications}, month = {3}, pages = {129660}, shortjournal = {Expert Syst. Appl.}, title = {SH-ETRs:Learning soft-hard rules with entity type constraints for document-level relation extraction}, volume = {298}, year = {2026}, }
• (2026). MCAD-EUC: Multi-context adaptive decoding with entropy-based uncertainty calibration for knowledge conflict mitigation. ESWA, 298, 129659. (www)

  @article{ , abstract = {The knowledge sources of large language models (LLMs) encompass both parametric internal knowledge and external contextual information. However, conflicts between these two sources can significantly impair model performance. Existing methods typically assume a priori correctness of either the context or the parametric knowledge, lacking dynamic coordination mechanisms and being limited to single-context scenarios. To address this issue, this work proposes a lightweight and training-free decoding method, M ulti- C ontext A daptive D ecoding ( MCAD-EUC ), which dynamically measures the effectiveness of both knowledge through E ntropy based U ncertainty C alibration. It does not concern itself with whether the knowledge is false or true, the internal or the external, but balancing them according to their contributions to correctly answering the question. Particularly, MCAD-EUC is naturally multi-contextual. It can dynamically amplify the distribution of golden context while mitigating the influence of noisy context, thereby optimizing the final logits for predicting the next token during the decoding process. To comprehensively evaluate the model performance in multi-context scenarios, this work constructs MCQA, a multi-context question answering dataset that includes golden context, irrelevant context, and six categories of misleading context (crowd, logic, temporal, authority, emotional, numeric), simulating the diversity of noise in real-world settings. Extensive experiments on four LLMs and four MCQA datasets demonstrate that MCAD-EUC achieves an average accuracy improvement of 3.17 % over the best-performing baseline methods. Further sensitivity analysis confirms that the entropy-based adaptive weighting mechanism consistently outperforms all fixed-weight settings. Our dataset and code will be publicly available.}, archive = {J_ESWA}, author = {Yimin Ou and Yifan Wang and Ping Jian and Tianhe Zhang and Xing Pei}, doi = {10.1016/j.eswa.2025.129659}, journal = {Expert Systems with Applications}, month = {3}, pages = {129659}, shortjournal = {Expert Syst. Appl.}, title = {MCAD-EUC: Multi-context adaptive decoding with entropy-based uncertainty calibration for knowledge conflict mitigation}, volume = {298}, year = {2026}, }
• (2026). Interpretable knowledge tracing with dual-level knowledge states. ESWA, 298, 129658. (www)

  @article{ , abstract = {Knowledge Tracing (KT) is a critical technology for achieving personalized learning. It estimates learners’ knowledge states and predicts future performance using historical interaction data. Despite recent advances, two significant challenges remain. First, the accuracy of knowledge state modeling is limited by the insufficient fusion of multi-scale information across temporal and spatial dimensions. Second, a trade-off persists between improving predictive performance and enhancing interpretability. This paper proposes an interpretable knowledge tracing method based on dual-level knowledge states (DIKT) to address these challenges. From the temporal perspective, DIKT incorporates a forgetting-aware RoLinear Transformer and a semantic similarity-based review mechanism to model learners’ problem-level knowledge states. From the spatial perspective, it leverages a Knowledge Concept (KC) relational graph to propagate influence among related KCs and dynamically update learners’ concept-level knowledge states through three sequential learning phases: forgetting, aggregation, and updating. Student performance is predicted using a two-parameter Item Response Theory (IRT) model, which incorporates guess and slip parameters to account for response anomalies. We conduct extensive comparisons between DIKT and 20 state-of-the-art KT models on five widely used public datasets. Experimental results demonstrate that DIKT achieves superior performance while preserving interpretability, highlighting its practical potential for real-world educational applications. The code is available at https://github.com/ting214/DIKT .}, archive = {J_ESWA}, author = {Yanting Li and Tao Zhou and Tianyu Cai and Shenggen Ju}, doi = {10.1016/j.eswa.2025.129658}, journal = {Expert Systems with Applications}, month = {3}, pages = {129658}, shortjournal = {Expert Syst. Appl.}, title = {Interpretable knowledge tracing with dual-level knowledge states}, volume = {298}, year = {2026}, }
• (2026). RASpan: Improving toponym recognition through span representation model with retrieval augmentation. ESWA, 298, 129657. (www)

  @article{ , abstract = {Toponym recognition aims to identify place names from natural language texts, which is vital for various applications including geographic information retrieval, emergency response, and natural disaster analysis. Currently, mainstream studies mainly adopt deep learning models for toponym recognition. However, these approaches encounter significant limitations due to the inherent ambiguity, variation, and abbreviation of toponyms. To address these issues, we propose a novel Span Representation Model with R etrieval A ugmentation ( RASpan ) that leverages more accurate span representation and effective external geo-entity information to enhance the semantic representation of place names for improving the performance of toponym recognition. On the one hand, RASpan retrieves diverse geo-entities and concatenates geo-entity knowledge with an input sequence to construct a new prompt sequence. On the other hand, RASpan utilizes the prompt encoder based on the language model to encode this prompt sequence and employs a dedicated span representation module to obtain more accurate span representations. In addition, a new geo-entity prediction task is designed to learn the entire representation of each geo-entity while minimizing noise interference. Experiments on three publicly available datasets demonstrate that our model achieves new state-of-the-art results, highlighting the effectiveness of RASpan in toponym recognition by introducing prior geo-entity knowledge.}, archive = {J_ESWA}, author = {Hui Wu and Anran Yang and Zhinong Zhong and Ye Wu and Fei Yang and Luo Chen and Ning Jing}, doi = {10.1016/j.eswa.2025.129657}, journal = {Expert Systems with Applications}, month = {3}, pages = {129657}, shortjournal = {Expert Syst. Appl.}, title = {RASpan: Improving toponym recognition through span representation model with retrieval augmentation}, volume = {298}, year = {2026}, }
• (2026). An interpretable automated optimized machine learning for predicting concrete compressive strength. ESWA, 298, 129656. (www)

  @article{ , abstract = {This study presents a novel, interpretable, and automated machine learning (AutoML) framework for accurately predicting the compressive strength of environmentally sustainable concrete mixtures that incorporate supplementary cementitious materials (SCMs) by addressing the growing need for transparent and data driven tools in structural material design, particularly for concrete mixes enriched with various SCMs. A robust unified dataset of 1,317 samples was curated by integrating peer-reviewed experimental studies for this study. The proposed methodology incorporates feature contribution ranking through mutual information, model screening with AutoML to identify the most effective regression models, Bayesian optimization for fine-tuning model parameters, and interpretability techniques including SHAP and counterfactual analysis. The best performance metrics, in training include R 2 of 0.999, a mean absolute error 0.114, root mean squared error 0.7094, and mean absolute percentage error of 0.51 %, in testing phase R 2 of 0.944, a mean absolute error 3.479, root mean squared error 4.8173, and mean absolute percentage error of 9.86 %. The weakest performance, with a training R 2 of 0.982 and a mean absolute error of 1.911 MPa, root mean squared error 2.7990 and mean absolute percentage error 5.43 %, in testing phase R 2 of 0.786 and a mean absolute error of 5.754 MPa, root mean squared error 9.4336 and mean absolute percentage error 16.16 %. The interpretability analysis values provided insights into the most important features, such as curing time and cement are crucial in predicting the strength. Counterfactual analysis further validated the model by illustrating the significant impact of cement, age and water on concrete strength.}, archive = {J_ESWA}, author = {Aparna Kamarthi and Baskar Kaliyamoorthy}, doi = {10.1016/j.eswa.2025.129656}, journal = {Expert Systems with Applications}, month = {3}, pages = {129656}, shortjournal = {Expert Syst. Appl.}, title = {An interpretable automated optimized machine learning for predicting concrete compressive strength}, volume = {298}, year = {2026}, }
• (2026). The power of text similarity in identifying AI-LLM paraphrased documents: The case of BBC news articles and ChatGPT. ESWA, 298, 129655. (www)

  @article{ , abstract = {Generative AI paraphrased text can be used for copyright infringement and the AI paraphrased content can deprive substantial revenue from original content creators. Despite this recent surge of malicious use of generative AI, there are few academic publications that research this threat. In this article, we demonstrate the ability of pattern-based similarity detection for AI paraphrased news recognition. We propose an algorithmic scheme, which is not limited to detect whether an article is an AI paraphrase, but, more importantly, to identify that the source of infringement is the ChatGPT. The proposed method is tested with a benchmark dataset specifically created for this task that incorporates real articles from BBC, incorporating a total of 2,224 articles across five different news categories, as well as 2,224 paraphrased articles created with ChatGPT. Results show that our pattern similarity-based method, that makes no use of deep learning, can detect ChatGPT assisted paraphrased articles at percentages 96.23% for accuracy, 96.25 for precision, 96.21% for sensitivity, 96.25% for specificity and 96.23% for F1 statistic.}, archive = {J_ESWA}, author = {Konstantinos F. Xylogiannopoulos and Petros Xanthopoulos and Panagiotis Karampelas and Georgios A. Bakamitsos}, doi = {10.1016/j.eswa.2025.129655}, journal = {Expert Systems with Applications}, month = {3}, pages = {129655}, shortjournal = {Expert Syst. Appl.}, title = {The power of text similarity in identifying AI-LLM paraphrased documents: The case of BBC news articles and ChatGPT}, volume = {298}, year = {2026}, }
• (2026). Dynamic collaborative evolutionary network: A novel spatio-temporal feature extraction framework for EEG emotion recognition. ESWA, 298, 129654. (www)

  @article{ , abstract = {In recent years, significant progress has been made in emotion recognition research based on electroencephalogram (EEG) signals. However, existing methods face two key limitations: on one hand, the reliance on fixed physical connections or static topological relationships makes it difficult to effectively represent the dynamic non-Euclidean spatial characteristics between EEG electrodes; on the other hand, spatiotemporal feature extraction is often conducted independently. This lack of a collaborative mechanism for spatiotemporal features results in insufficient fine-grained emotional representation capability. To address these issues, a dynamic collaborative evolutionary network (DCENet) is proposed based on graph-aware enhancement and global convolutional Transformer for EEG emotion recognition. DCENet constructs the causal relationship between electrodes by constructing the graph-aware enhancement (GAE) module, obtains spatial features with the causal relationship, and enhances key features. At the same time, DCENet constructs the global convolutional Transformer (GCT) module, which utilizes the global modeling advantage of the Transformer and the local perception ability of the convolutional operation to capture the temporal features with different scales. In addition, DCENet adaptively fuses temporal and spatial features through the local differential fusion (LDF) module to achieve cross-domain feature alignment and feature alignment of emotion categories to collaboratively evolve emotion representation features with more fine-grained information. This paper conducts experiments on the SEED, SEED-IV, and MPED datasets to validate the effectiveness of DCENet. The experimental results show that the model achieves cross-subject average accuracies of 87.55 %, 73.04 %, and 27.72 % on SEED, SEED-IV, and MPED, respectively, outperforming the state-of-the-art methods. The source code is publicly available at: https://github.com/cvmdsp/DCENet .}, archive = {J_ESWA}, author = {Shuaiqi Liu and Zhihui Gu and Yuan Zhang and Yanling An and Shuhuan Zhao and Bing Li and Yudong Zhang}, doi = {10.1016/j.eswa.2025.129654}, journal = {Expert Systems with Applications}, month = {3}, pages = {129654}, shortjournal = {Expert Syst. Appl.}, title = {Dynamic collaborative evolutionary network: A novel spatio-temporal feature extraction framework for EEG emotion recognition}, volume = {298}, year = {2026}, }
• (2026). Trust in recommender systems: A survey. ESWA, 298, 129653. (www)

  @article{ , abstract = {Trust-based recommender systems incorporate interpersonal trust relationships into the recommendation process, operating on the principle that users are more likely to accept suggestions from people they trust. Empirical studies have shown that trust-aware approaches often deliver more accurate recommendations than their trust-unaware counterparts. In this comprehensive, up-to-date survey, we analyze a variety of trust-based recommendation methods, categorize different trust inference techniques, and examine how trust is integrated into recommendation algorithms. We then organize the investigated approaches according to a clear taxonomy, explore their underlying concepts, and highlight the key challenges and open issues that remain in the field.}, archive = {J_ESWA}, author = {Imane Akdim and Loubna Mekouar and Youssef Iraqi}, doi = {10.1016/j.eswa.2025.129653}, journal = {Expert Systems with Applications}, month = {3}, pages = {129653}, shortjournal = {Expert Syst. Appl.}, title = {Trust in recommender systems: A survey}, volume = {298}, year = {2026}, }
• (2026). Composite human activity recognition utilizing knowledge distillation and sensor fusion focusing on resource constrained microcontrollers. ESWA, 298, 129652. (www)

  @article{ , abstract = {This study presents a cost-effective, low-computation system for composite Human Activity Recognition (HAR) that leverages knowledge-distilled neural networks on a Microcontroller Unit (MCU) to minimize reliance on cloud processing. A key contribution of this work is the investigation of plantar pressure sensor data within a knowledge distillation framework, addressing a notable gap in the existing literature. The proposed solution centers around the ESP32-S3 DevKit C1, equipped with a dual-core 240 MHz Tensilica chip, 320 KiB of usable Static Random Access Memory (SRAM), and built-in Wi-Fi and Bluetooth. Significantly, both the teacher and the student models surpass existing state-of-the-art methods, achieving F1-scores of 99.33 %, 98.36 %, and 97.68 % respectively, in classifying a comprehensive set of 21 activities (15 composite and 6 simple). The distilled student models demonstrate remarkable efficiency, with execution times of 1.83 and 0.64 s, memory footprints of only 62 KB and 82 KB, and flash memory usage of approximately 209 KB and 127 KB, while maintaining low power consumption of 210 mW and 215 mW, respectively. Furthermore, we have developed an end-to-end prototype that integrates the ESP32-S3 with a WitMotion Inertial Measurement Unit (IMU) sensor. This system autonomously manages data acquisition, feature extraction, and inference in under 7 s with a total power consumption of approximately 295 mW.}, archive = {J_ESWA}, author = {Athar Noor Mohammad Rafee and John Clear and Jannatun Noor}, doi = {10.1016/j.eswa.2025.129652}, journal = {Expert Systems with Applications}, month = {3}, pages = {129652}, shortjournal = {Expert Syst. Appl.}, title = {Composite human activity recognition utilizing knowledge distillation and sensor fusion focusing on resource constrained microcontrollers}, volume = {298}, year = {2026}, }
• (2026). Med-D3CG: Wavelet-based diffusion in the difference domain for cross-modality medical image generation. ESWA, 298, 129651. (www)

  @article{ , abstract = {The synthesis of cross-modal medical images plays a vital role in bridging diagnostic gaps between imaging modalities such as CT, MRI, and PET. This integration enables a more comprehensive evaluation of a patient’s condition, improving diagnostic accuracy and aiding clinical decision-making. However, the performance of conditional denoising diffusion probabilistic models is often hindered by pronounced structural and intensity discrepancies between modalities, as well as the inherently slow nature of the diffusion process. To address these challenges, this paper proposes Wavelet-Based Diffusion in the Difference Domain for Cross-Modality Medical Image Generation (Med-D3CG), a novel framework that transforms the synthesis process by emphasizing the difference domain. Instead of directly generating target images like conventional methods, Med-D3CG models the residual information between conditioned and target images. This strategy allows the framework to accurately capture essential structural and intensity variations between modalities, leading to more precise and realistic image synthesis. Additionally, Med-D3CG integrates the Discrete Wavelet Transform (DWT) to improve efficiency, accelerating the diffusion process while maintaining high image fidelity. On SynthRAD2023 and HMIFD datasets, state-of-the-art performance is achieved on pelvis and HMIFD using Med-D3CG , with the best Learned Perceptual Image Patch Similarity (LPIPS) and competitive FID observed on brain. Code and pretrained models are provided at https://github.com/ZgzTTTer/Med-D3CG .}, archive = {J_ESWA}, author = {Guangzhen Zhu and Midi Wan and Wenming Cao and Zhiwen Yu and Jin Hu and Bing Li and Xiaotao Fan}, doi = {10.1016/j.eswa.2025.129651}, journal = {Expert Systems with Applications}, month = {3}, pages = {129651}, shortjournal = {Expert Syst. Appl.}, title = {Med-D3CG: Wavelet-based diffusion in the difference domain for cross-modality medical image generation}, volume = {298}, year = {2026}, }
• (2026). Two-stage feature selection utilizing three-way adaptive neighborhood characteristic measure and optimal combination search. ESWA, 298, 129650. (www)

  @article{ , abstract = {The neighborhood rough set model(NRSM) has shown its powerful capacity in feature selection. However, a challenge still exists in describing the diversity between the attributes deeply while avoiding the impact caused by the neighborhood parameters. To address this problem, in this paper, we propose a two-stage feature selection by utilizing a three-way adaptive characteristic measure and an optimal combination search. First, we define a fitness function for applying the Stochastic Fractal Search(SFS) to design an novel adaptive neighborhood rough set model(ANRSM). To better utilize the construction characteristic of the adaptive model and reduce the computational cost, the lower and upper approximations of the SFS-based ANRSM are redefined through the fitness function. Second, based on the two approximations, we analyze the fitness function thresholds that can partition the universe into three regions and design the three-way adaptive neighborhood characteristic regions, which provide a more direct classification of samples without the inclusion and union operations. Third, we design different measures for the samples in diverse regions based on their corresponding characteristic. Afterward, a three-way adaptive characteristic measure is designed by integrating the three measures to evaluate the uncertainty of attributes. Then, we apply the measure to design a feature selection approach with greedy search. Considering that the greedy strategy may output redundant attributes, we introduce an optimal combination search approach through a novel wrapper technology to explore the potential optimal feature combinations. Compared with nine algorithms on fourteen public datasets, the experimental results show the effectiveness of our algorithm.}, archive = {J_ESWA}, author = {Bowen Lin and Duoqian Miao and Caihui Liu and Hongyun Zhang and Ruizhi Wang and Witold Pedrycz}, doi = {10.1016/j.eswa.2025.129650}, journal = {Expert Systems with Applications}, month = {3}, pages = {129650}, shortjournal = {Expert Syst. Appl.}, title = {Two-stage feature selection utilizing three-way adaptive neighborhood characteristic measure and optimal combination search}, volume = {298}, year = {2026}, }
• (2026). GE-adapter: A general and efficient adapter for enhanced video editing with pretrained text-to-image diffusion models. ESWA, 298, 129649. (www)

  @article{ , abstract = {Recent advancements in text-to-image (T2I) generation using diffusion models have enabled cost-effective video-editing applications by leveraging pre-trained models, eliminating the need for resource-intensive training. However, the frame-independence of T2I generation often results in poor temporal consistency. Existing methods address this issue through temporal layer fine-tuning or inference-based temporal propagation, but these approaches suffer from high training costs or limited temporal coherence. To address these challenges, we propose a General and Efficient Adapter (GE-Adapter) that integrates temporal-spatial and semantic consistency with Baliteral DDIM inversion. This framework introduces three key components: (1) Frame-based Temporal Consistency Blocks (FTC Blocks) to capture frame-specific features and enforce smooth inter-frame transitions via temporally-aware loss functions; (2) Channel-dependent Spatial Consistency Blocks (SCD Blocks) employing bilateral filters to enhance spatial coherence by reducing noise and artifacts; and (3) Token-based Semantic Consistency Module (TSC Module) to maintain semantic alignment using shared prompt tokens and frame-specific tokens. Our method significantly improves perceptual quality, text-image alignment, and temporal coherence, as demonstrated on the MSR-VTT dataset. Additionally, it achieves enhanced fidelity and frame-to-frame coherence, offering a practical solution for T2V editing. The project page is https://github.com/codepassionor/T2I_Adapter .}, archive = {J_ESWA}, author = {Yangfan He and Sida Li and Kun Li and Jianhui Wang and Binxu Li and Tianyu Shi and Yi Xin and Keqin Li and Jun Yin and Miao Zhang and Xueqian Wang}, doi = {10.1016/j.eswa.2025.129649}, journal = {Expert Systems with Applications}, month = {3}, pages = {129649}, shortjournal = {Expert Syst. Appl.}, title = {GE-adapter: A general and efficient adapter for enhanced video editing with pretrained text-to-image diffusion models}, volume = {298}, year = {2026}, }
• (2026). New semi-supervised fuzzy C-means clustering with asymmetric deviation constraints and fast algorithm. ESWA, 298, 129648. (www)

  @article{ , abstract = {Semi-supervised clustering leverages prior information to improve algorithm performance and is widely valued by researchers. This paper analyzes the traditional semi-supervised fuzzy C-means (SFCM) objective function, noting that as a labeled sample’s membership degree aligns with its prior information, the impact of this information on the deviation constraint weakens. This reduces its supervisory effect on optimizing the membership partition matrix, especially with a large regularization factor. To overcome this, we propose a novel semi-supervised fuzzy C-means method based on an asymmetric deviation constraint and develop a two-level alternating iterative optimization algorithm, supported by theoretical convergence analysis using Zangwill’s theorem and the bordered Hessian matrix. To address the slow convergence and high computational cost typical of semi-supervised fuzzy clustering, we further enhance the algorithm with affinity filtering and a membership scaling scheme for improved efficiency. Experimental results demonstrate that our methods significantly outperform existing state-of-the-art techniques, advancing semi-supervised fuzzy C-means clustering.}, archive = {J_ESWA}, author = {Chengmao Wu and Jun Hou}, doi = {10.1016/j.eswa.2025.129648}, journal = {Expert Systems with Applications}, month = {3}, pages = {129648}, shortjournal = {Expert Syst. Appl.}, title = {New semi-supervised fuzzy C-means clustering with asymmetric deviation constraints and fast algorithm}, volume = {298}, year = {2026}, }
• (2026). BayesAHDD: A new bayesian rule-based adaptive hypersphere data description for few-shot one-class classification. ESWA, 298, 129647. (www)

  @article{ , abstract = {Few-shot one-class classification (FS-OCC) is a challenging classification problem that involves learning from a very limited number of training samples, all from a single class. Recently, several data description methods have been proposed to address the FS-OCC problem. Unlike conventional one-class classification problems, the few-shot setting requires the model to generalize to novel tasks with previously unseen positive classes. Most existing methods learn decision boundaries in the feature space without explicitly modeling the underlying data distributions, which limits the generalization ability of the learned representations. To address this issue, we propose Bayesian Rule-based Adaptive Hypersphere Data Description (BayesAHDD), a probabilistic framework that represents data with multivariate Gaussian distributions and performs classification according to the Bayes decision rule. Based on the assumption that negative samples are more dispersed in the feature space, BayesAHDD models the negative class by scaling the positive class variance vector element-wise using a learnable vector. To address the challenges of exploding gradients and numerical overflow, we impose a lower bound on the positive class variance vector and introduce a trainable parameter that integrates the class prior probability ratio with the normalization constants of the Gaussian class-conditional densities. Experimental results on both benchmark and domain-specific datasets show that BayesAHDD consistently outperforms existing baselines and state-of-the-art FS-OCC methods. Moreover, quantitative analysis demonstrates that the learned feature representations exhibit superior discriminative ability compared to those produced by previous approaches.}, archive = {J_ESWA}, author = {Yuchen Ren and Xiabi Liu and Yan Pei and Yunlong Li and Yongxia Wei}, doi = {10.1016/j.eswa.2025.129647}, journal = {Expert Systems with Applications}, month = {3}, pages = {129647}, shortjournal = {Expert Syst. Appl.}, title = {BayesAHDD: A new bayesian rule-based adaptive hypersphere data description for few-shot one-class classification}, volume = {298}, year = {2026}, }
• (2026). A network selection algorithm for space-air-ground integrated network based on location prediction model and multi-attribute decision making. ESWA, 298, 129646. (www)

  @article{ , abstract = {As an indispensable component of 5G and even the future 6G networks, the Space-Air-Ground Integrated Network (SAGIN) is envisioned to provide ubiquitous network connectivity and services by integrating satellite, aerial, and terrestrial networks. However, due to the frequent network selection of in-vehicle terminals, the user’s Quality of Service (QoS) can significantly deteriorate. To address this issue, a network selection algorithm based on terminal location prediction has been proposed. Firstly, we enhanced the Particle Swarm Optimization (PSO) algorithm to optimize the hyper-parameters of the Long Short-Term Memory (LSTM) network, thereby improving the accuracy of terminal location prediction. After constructing the network sets of the current terminal position and the predicted position, respectively, we designed a network selection judgment mechanism with a dynamically adjustable switching threshold based on Fuzzy Logic and K-Means theory. Finally, through the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) algorithm, we have achieved robust network selection in fast-moving scenarios. The simulation results show that the proposed algorithm can adaptively adjust the switching threshold and provide precise positions. Compared to existing algorithms, it can significantly reduce the number of candidate networks and the number of selections, thereby reducing the computational load and increasing the throughput of users.}, archive = {J_ESWA}, author = {Jianli Xie and Weicheng Pan and Lei Wu and Zishan Wu}, doi = {10.1016/j.eswa.2025.129646}, journal = {Expert Systems with Applications}, month = {3}, pages = {129646}, shortjournal = {Expert Syst. Appl.}, title = {A network selection algorithm for space-air-ground integrated network based on location prediction model and multi-attribute decision making}, volume = {298}, year = {2026}, }
• (2026). Anomaly detection based on graph neural networks incorporating with domain knowledge for industrial cyber-physical systems. ESWA, 298, 129645. (www)

  @article{ , abstract = {Effective anomaly detection heavily relies on accurately modeling the normal behavior of the industrial cyber-physical systems (ICPSs). Current popular data-driven black-box methods’ performance depends on the quantity and quality of data from the ICPS, rather than integrating with the intrinsic characteristics of the system, such as mass conservation and structural dependencies of industrial processes. The data used to train these models often fails to cover all operating conditions of ICPSs, resulting in a distributional mismatch between training and testing data. This limitation significantly reduces the generalization capability of trained data-driven models for detecting anomalies under unseen and unknown operational conditions. Meanwhile, constructing well-generalized behavior models using white-box mechanism models (MM)–which represent the internal behavior of controllers and physical processes using first-principles equations or physical laws, such as conservation principles and kinetic dynamics–is not always feasible in real-world industrial applications. To address this issue, this paper presents a novel anomaly detection approach employing a domain knowledge-embedded hybrid graph neural network (HGNN) that integrates control, physical, and structural characteristics of ICPSs into a data-driven modeling framework. It establishes a domain knowledge-based MM to predict the behavior of controllers and physical processes, and utilizes a HGNN model with embedded state topology and spatial distribution knowledge to compensate for the predictive error of MM, forming a hybrid model called mechanism-embedded HGNN (MEHGNN). This model can detect anomalies in unknown operational conditions by analyzing the predictive error of MEHGNN. In the experiments, MEHGNN raises the detection accuracy of GNNs from 36.6 %–82.4 % to 79.4 %–96.3 % under distribution shift scenarios, achieving the best detection performance among all compared methods.}, archive = {J_ESWA}, author = {Xin Du and Chunjie Zhou and Yu-Chu Tian and Bo Xu}, doi = {10.1016/j.eswa.2025.129645}, journal = {Expert Systems with Applications}, month = {3}, pages = {129645}, shortjournal = {Expert Syst. Appl.}, title = {Anomaly detection based on graph neural networks incorporating with domain knowledge for industrial cyber-physical systems}, volume = {298}, year = {2026}, }
• (2026). Predictive modeling and multi-objective optimization of screw whirling milling based on ISSA-BP embedded NSGA-III algorithm. ESWA, 298, 129644. (www)

  @article{ , abstract = {In screw whirling milling, the relationship between machining quality and processing parameters exhibits highly nonlinear characteristics. The traditional multiple regression models may not be able to capture this complex relationship accurately. Therefore, it is necessary to consider more flexible and applicable algorithms to establish their connections and optimize processing parameters. It can improve the accuracy and reliability of products, and provide more scientific method guidance for screw whirling milling processing. The originality of this article lies in proposing an adaptive dynamic optimization hybrid model. This model combines improved sparrow search algorithm optimized backpropagation (ISSA-BP) and non-dominated sorting genetic algorithm (NSGA-III). It can effectively adapt to dynamic data and find the optimal balance point among multiple objectives to better predict and optimize responses (cutting force, vibration, roughness, and residual compressive stress) in screw whirling milling. Firstly, a suitable network structure is identified by comparing the effects of five improvement strategies, population size, and the ratio of producers to scouters on the sparrow search algorithm. Then, an ISSA-BP prediction model is developed for four responses based on this structure. On this basis, the superiority of the established ISSA-BP model is verified by comparing prediction performance of five algorithms, and the relative prediction errors are all within 2%. The R 2 values of the models are all above 0.99, and they also perform well in indicators such as MAE (Mean Absolute Error), MSE (Mean Squared Error), MAPE (Mean Absolute Percentage Error), and RMSE (Root Mean Squared Error). Then, ISSA-BP model is encapsulated and embedded into the optimization algorithm as the fitness function of NSGA-III. Finally, with the processing parameters of whirling milling as constraints, the NSGA-III algorithm is used to solve the proposed model and obtain the Pareto optimal solution set. Choosing appropriate processing parameters according to different needs in actual machining can help improve the quality and efficiency of screw machining.}, archive = {J_ESWA}, author = {Chao Liu and Hao Ding and Juanjuan Zheng and Yan He and Shaofu Huang and Junbo Tuo and Zuqing Luo and Gang Shen}, doi = {10.1016/j.eswa.2025.129644}, journal = {Expert Systems with Applications}, month = {3}, pages = {129644}, shortjournal = {Expert Syst. Appl.}, title = {Predictive modeling and multi-objective optimization of screw whirling milling based on ISSA-BP embedded NSGA-III algorithm}, volume = {298}, year = {2026}, }
• (2026). A survey of large language models for data challenges in graphs. ESWA, 298, 129643. (www)

  @article{ , abstract = {Graphs are a widely used paradigm for representing non-Euclidean data, with applications ranging from social network analysis to biomolecular prediction. While graph learning has achieved remarkable progress, real-world graph data presents a number of challenges that significantly hinder the learning process. In this survey, we focus on four fundamental data-centric challenges: (1) Incompleteness , real-world graphs have missing nodes, edges, or attributes; (2) Imbalance , the distribution of the labels of nodes or edges and their structures for real-world graphs are highly skewed; (3) Cross-domain Heterogeneity , graphs from different domains exhibit incompatible feature spaces or structural patterns; and (4) Dynamic Instability , graphs evolve over time in unpredictable ways. Recently, Large Language Models (LLMs) offer the potential to tackle these challenges by leveraging rich semantic reasoning and external knowledge. This survey focuses on how LLMs can address four fundamental data-centric challenges in graph-structured data, thereby improving the effectiveness of graph learning. For each challenge, we review both traditional solutions and modern LLM-driven approaches, highlighting how LLMs contribute unique advantages. Finally, we discuss open research questions and promising future directions in this emerging interdisciplinary field. To support further exploration, we have curated a repository of recent advances on graph learning challenges: https://github.com/limengran98/Awesome-Literature-Graph-Learning-Challenges .}, archive = {J_ESWA}, author = {Mengran Li and Pengyu Zhang and Wenbin Xing and Yijia Zheng and Klim Zaporojets and Junzhou Chen and Ronghui Zhang and Yong Zhang and Siyuan Gong and Jia Hu and Xiaolei Ma and Zhiyuan Liu and Paul Groth and Marcel Worring}, doi = {10.1016/j.eswa.2025.129643}, journal = {Expert Systems with Applications}, month = {3}, pages = {129643}, shortjournal = {Expert Syst. Appl.}, title = {A survey of large language models for data challenges in graphs}, volume = {298}, year = {2026}, }
• (2026). Region-aware prediction strategy based on shared points and multiple scales for dynamic multi-objective optimization. ESWA, 298, 129642. (www)

  @article{ , abstract = {In dynamic multi-objective optimization problems, effectively predicting and tracking the Pareto optimal front (POF) under environmental changes has been one of the core challenges. In this paper, we propose a region-aware prediction strategy based on shared points and multiple scales (RADMOEA) that combines global and local characteristics, aiming to enhance the algorithm’s ability to sense and adapt to POF. Firstly, the center-point movement strategy is used to move the non-dominated solution set from the previous moment to obtain the non-dominated solution set after the movement. The actual non-dominated solution set at the current moment and the non-dominated solution set after the movement share points in the objective space, and these shared points divide the non-dominated solution set at the current moment into several subregions. Within each region, all individuals are appropriately rescaled, and a local coordinate system is established. Then, within the local coordinate system, each individual is associated with the nearest post-movement non-dominated individual. Finally, new populations adapted to environmental changes are generated by combining centroid movement directions, Gaussian perturbations, and multi-scale individual association relationships. The proposed strategy is compared with six advanced algorithms, and the experimental results demonstrate that RADMOEA is effective in tracking the POF under dynamic environments.}, archive = {J_ESWA}, author = {Yaru Hu and Sitong Wang and Junwei Ou and Zhenlin Mei and Juan Zou and Shengxiang Yang}, doi = {10.1016/j.eswa.2025.129642}, journal = {Expert Systems with Applications}, month = {3}, pages = {129642}, shortjournal = {Expert Syst. Appl.}, title = {Region-aware prediction strategy based on shared points and multiple scales for dynamic multi-objective optimization}, volume = {298}, year = {2026}, }
• (2026). A transformer network for multi-dimensional nonuniform aperture synthesis radiometer image inversion. ESWA, 298, 129641. (www)

  @article{ , abstract = {Passive microwave remote sensing plays a vital role in Earth observation, with applications in soil moisture, ocean salinity, and atmospheric monitoring. However, improving spatial resolution at low frequencies remains challenging. Recently, combining multiple small antenna arrays into a larger one has emerged as a technological approach to enhance spatial resolution. Nevertheless, aperture synthetic radiometers formed by such combinations usually consist of non-uniform antenna arrays (one-dimensional, two-dimensional, or three-dimensional). Compared with regular antenna arrays, they complicate the inversion of brightness temperature (BT) images. This paper proposes NASRT, a transformer-based inversion method designed for multi-dimensional non-uniform ASRs. The network extracts and fuses spectral and UVW spatial distribution features from the visibility function (VF), and introduces a learnable position weight matrix during training to capture spatial information of the non-uniform array. Through supervised learning, NASRT effectively maps the VF to BT images. Simulations across 1-D, 2-D, and 3-D NASR scenes demonstrate that NASRT achieves higher accuracy and stability than traditional methods. In a 1-D NASR indoor experiment, the proposed method also shows improved inversion accuracy and lower sidelobes, validating its effectiveness.}, archive = {J_ESWA}, author = {Jian Dong and Jiaxin Li and Chengwang Xiao and Rigeng Wu and Haofeng Dou and Wenjing Wang and Yuanchao Wu and Liangbing Chen}, doi = {10.1016/j.eswa.2025.129641}, journal = {Expert Systems with Applications}, month = {3}, pages = {129641}, shortjournal = {Expert Syst. Appl.}, title = {A transformer network for multi-dimensional nonuniform aperture synthesis radiometer image inversion}, volume = {298}, year = {2026}, }
• (2026). OceanAgent: A small-scale multi-modal assistant for ocean exploration. ESWA, 298, 129640. (www)

  @article{ , abstract = {Extraction of key information and the subsequent generation of actionable knowledge from multimodal data are critical for ocean exploration. Traditional knowledge generation methods rely heavily on expert experience and are labor-intensive. Recently, Large Multimodal Models (LMMs) have shown exceptional capabilities for knowledge generation from multimodal data in many complex tasks. These models also have potential to assist knowledge mining in ocean exploration. However, two major challenges faced by the LMMs when used in ocean exploration include the scarcity of ocean instruction-following data and the degradation of underwater visual environments. In this paper, a small-scale LMM for ocean exploration, named the OceanAgent, is designed. First, a swarm-intelligence-based leaderless multi-agent collaboration framework is proposed to generate visual instruction-following data. Subsequently, we present a visual-language connector to simultaneously extract multi-scale features. It is formed by integrating a multi-scale residual network with a multi-layer perceptron, which can enhance the model’s performance on severely low-quality images. Experiments show that the proposed method for constructing visual instruction-following datasets improves both the textual quality and visual dialogue. When severely degraded ocean visual data are processed using the trained OceanAgent, the image description accuracy and image comprehension are improved by 23.6 % and 21.6 %, respectively, compared to existing models. Additionally, the model demonstrates superior domain expertise, with a 95.7 % win rate in dialogue quality assessments.}, archive = {J_ESWA}, author = {Yun Xu and Yue Liu and Junpeng Shang and Jianmin Lin and Dongfang Ma}, doi = {10.1016/j.eswa.2025.129640}, journal = {Expert Systems with Applications}, month = {3}, pages = {129640}, shortjournal = {Expert Syst. Appl.}, title = {OceanAgent: A small-scale multi-modal assistant for ocean exploration}, volume = {298}, year = {2026}, }
• (2026). SPS-GAD: Spectral-spatial graph structure learning for anomaly detection in heterophilic graphs. ESWA, 298, 129639. (www)

  @article{ , abstract = {Graph anomaly detection(GAD) plays a critical role in fields such as fraud detection and network security. Although existing graph anomaly detection methods have achieved promising performance, most graph neural networks (GNNs) rely on the homophily assumption, which presumes that connected nodes share similar labels. However, real-world graphs frequently exhibit pronounced heterophily. Owing to class imbalance, normal nodes tend to have lower heterophily while anomalous nodes display higher heterophily. Furthermore, feature inconsistency induced by node camouflage exacerbates the detection challenge, rendering many existing approaches ineffective. To overcome these limitations, we propose SPS-GAD, a spectral-spatial graph structure learning framework specifically designed for detecting anomalous nodes in heterophilic graphs. First, to alleviate the feature inconsistency resulting from node camouflage, we develop a node reconstruction module that learns intermediate node representations to mitigate camouflage-induced bias, and applies spectral filters to extract the graph’s inherent structural features. Second, to address the heterophily disparities arising from class imbalance, we introduce a subgraph-type-aware spectral filtering module that leverages edge scores generated by an edge partitioner to segregate the graph into homophilic, ambiguous, and heterophilic subgraphs. Distinct spectral filters are subsequently applied to capture features across various frequency bands. Additionally, we integrate a neighbor-type-aware graph attention module that employs edge scores within an attention mechanism to guide the feature aggregation process, thereby enhancing spatial representation learning. Experimental evaluations on six real-world datasets reveal that SPS-GAD significantly outperforms all baseline methods in key metrics such as F1-Macro and AUC, thereby confirming its effectiveness in graph anomaly detection. The source code is publicly available at https://github.com/cozy24/SPS-GAD .}, archive = {J_ESWA}, author = {Chen Zhu and Yaying Zhang}, doi = {10.1016/j.eswa.2025.129639}, journal = {Expert Systems with Applications}, month = {3}, pages = {129639}, shortjournal = {Expert Syst. Appl.}, title = {SPS-GAD: Spectral-spatial graph structure learning for anomaly detection in heterophilic graphs}, volume = {298}, year = {2026}, }
• (2026). An efficient vision transformer-based 3D adaptive residual densenet with gated recurrent unit for early detection of alzheimer disease from magnetic resonance imaging. ESWA, 298, 129638. (www)

  @article{ , abstract = {The precise identification of Alzheimer’s disease (AD) holds a crucial position in patient healthcare, particularly in the early stages. This early detection is essential, as it allows patients to understand the severity of the disease and take preventive measures before irreversible brain damage occurs. Although recent studies have harnessed machine learning techniques for Computer-Aided Diagnosis (CAD) of AD, many of these research efforts have encountered limitations in diagnostic performance. In this research, a novel AD detection system using Magnetic Resonance Imaging (MRI) is developed to identify subtle brain changes at an early stage through deep learning strategies. The goal is to reduce the risk for individuals facing challenges such as aging and brain-related diseases. The MRI images were obtained from the internet. The detection phase begins by receiving the MRI images. A Vision Transformer-based 3D Adaptive Residual DenseNet combined with a Gated Recurrent Unit (VARD-GRU) architecture is employed for effective AD identification. Here, the model parameters are optimized using the proposed Fitness-based Walrus Optimization Algorithm (FWOA). The AD detection system’s performance is evaluated against several existing frameworks using validation metrics to assess its effectiveness. The accuracy of the developed FWOA-VARD-GRU model is 94.11%. The results demonstrate that the FWOA-VARD-GRU model significantly enhances diagnostic accuracy, outperforming comparable methods. The model’s effectiveness confirms its strong capability in detecting early-stage AD. This system offers a promising approach to reduce risks for individuals affected by aging and neurological conditions by enabling timely and accurate detection.}, archive = {J_ESWA}, author = {Dr. A. Hemlathadhevi and Indumathy Paranthaman and Moorthy Agoramoorthy and Dr. Hari Kumar Palani}, doi = {10.1016/j.eswa.2025.129638}, journal = {Expert Systems with Applications}, month = {3}, pages = {129638}, shortjournal = {Expert Syst. Appl.}, title = {An efficient vision transformer-based 3D adaptive residual densenet with gated recurrent unit for early detection of alzheimer disease from magnetic resonance imaging}, volume = {298}, year = {2026}, }
• (2026). Pareto optimization of two-agent scheduling on parallel batch machines. ESWA, 298, 129637. (www)

  @article{ , abstract = {This paper considers a Pareto optimization problem of scheduling jobs of two competing agents on parallel batch machines. The jobs have equal processing times and non-identical job sizes. The objective is to find Pareto optimality and the corresponding schedules for minimizing both agents’ makespans. We analyze and identify an approximate Pareto region with a guarantee of 2-approximate Pareto optimal. We propose an integrated algorithm to find the approximate Pareto optimal points. Our computational study shows that the proposed algorithm outperforms the widely used non-dominated sorting genetic algorithm (NSGA-II), and that the obtained approximate Pareto optimal front is very close to the Pareto optimal front.}, archive = {J_ESWA}, author = {Cui-Lin Zhang and Guo-Qiang Fan}, doi = {10.1016/j.eswa.2025.129637}, journal = {Expert Systems with Applications}, month = {3}, pages = {129637}, shortjournal = {Expert Syst. Appl.}, title = {Pareto optimization of two-agent scheduling on parallel batch machines}, volume = {298}, year = {2026}, }
• (2026). Hybrid fleet composition and scheduling for road-based cross-border logistics under cost differentiation: A bi-level programming approach. ESWA, 298, 129636. (www)

  @article{ , abstract = {Under the dual pressure of explosive growth in cross-border e-commerce demand and increasing timeliness requirements from overseas customers, cross-border logistics service providers are compelled to establish logistics facilities and deploy fleets across multiple regions to ensure rapid response. However, during freight transportation, the lack of effective management over these complex and heterogeneous fleets—particularly in terms of fleet composition and routing decisions—has led to high transportation costs and low operational efficiency. This study is grounded in the practical operational context of cross-border logistics in the Guangdong–Hong Kong–Macau Greater Bay Area and models a multi-level, multi-node cross-border transportation network. To minimize the overall operational cost, the problem is addressed from two interrelated decision-making perspectives: fleet composition at the strategic level and routing planning at the operational level. Thus, a bi-level programming model is proposed to systematically capture the hierarchical structure and the logical relationship between these two decision layers. Furthermore, the model incorporates cost differences among trucks with different functional capabilities to reflect the significant disparity in logistics cost structures between domestic and overseas operations. To address the above multi-objective mixed-integer linear programming (MILP) problem, a tailored Non-dominated Sorting Genetic Algorithm II (MNSGA-II) is developed. Several key components of the algorithm are modified and enhanced to improve its search efficiency and solution quality in handling the problem’s complexity. Comparative experiments against classical algorithms demonstrate the superior solution quality and robustness of the proposed approach. The influence of cost differentials on composition and scheduling decisions is further analyzed, providing practical insights for the strategic planning of cross-border logistics systems.}, archive = {J_ESWA}, author = {Zhi Tang and Ting Qu and Yanghua Pan and George Q. Huang}, doi = {10.1016/j.eswa.2025.129636}, journal = {Expert Systems with Applications}, month = {3}, pages = {129636}, shortjournal = {Expert Syst. Appl.}, title = {Hybrid fleet composition and scheduling for road-based cross-border logistics under cost differentiation: A bi-level programming approach}, volume = {298}, year = {2026}, }
• (2026). A diversity-based niching differential evolution with neighborhood competition for nonlinear equation systems. ESWA, 298, 129635. (www)

  @article{ , abstract = {Solving nonlinear equation systems (NESs) has long been a fundamental challenge in the field of optimization. Due to the existence of multiple roots, such problems often exhibit complex and multimodal characteristics. Although numerous differential evolution-based algorithms have been developed to solve NESs, most of them employ only a single mutation operator, which is not adaptable to different problem scenarios. To this end, a diversity-based niching differential evolution with neighborhood competition (DNDE) is proposed to solve NESs. First, a control mechanism that takes into account population diversity and the evolutionary stage is proposed to adaptively assign appropriate mutation strategies to each subpopulation (niche), thereby enhancing the efficiency of root-finding. Second, a neighborhood priority competition mechanism is proposed to reduce cross-peak competition between populations, which ensures local convergence while improving global convergence. Finally, a reinitialization strategy based on opposition learning is introduced to guide the population toward more promising areas of the search space. Experimental results on 18 complex NESs and two real-world engineering problems show that DNDE outperforms many advanced algorithms in both root rate and success rate, demonstrating its effectiveness and value in practical applications.}, archive = {J_ESWA}, author = {Jianwei Li and Xinchao Zhao and Lingyu Wu and Yizhan Wu and Lingjuan Ye}, doi = {10.1016/j.eswa.2025.129635}, journal = {Expert Systems with Applications}, month = {3}, pages = {129635}, shortjournal = {Expert Syst. Appl.}, title = {A diversity-based niching differential evolution with neighborhood competition for nonlinear equation systems}, volume = {298}, year = {2026}, }
• (2026). Ill-posed regions-aware self-supervised stereo matching with left-right consistency. ESWA, 298, 129634. (www)

  @article{ , abstract = {Stereo matching presents a significant challenge due to the difficulty in accurately matching pixels between two images. These challenging pixels are typically found in ill-posed regions, which include areas with weak or repetitive textures, occlusions, and invisible regions. In recent years, there has been extensive research into stereo matching methods based on deep learning. However, these methods often struggle to accurately handle ill-posed regions, resulting in limited improvements in accuracy. To address this issue, this paper introduces a stereo matching model, IASSM-LRC, that is not constrained by disparity range. This model employs self-attention and cross-attention from the Transformer for dense attention computation and generates an initial disparity map. To effectively manage ill-posed regions, invalid masks are generated using left and right disparity maps. The initial disparity map is then fine-tuned through these invalid masks. IASSM-LRC is trained in an unsupervised manner. Experimental results from multiple different scenes demonstrate that the IASSM-LRC model exhibits strong disparity prediction performance. On KITTI2015, its D1-fg metric is 13.48 %, outperforming both PASM-Net’s 16.36 % and Flow2Stereo’s 14.62 %. The prediction results on the self-built calibration board dataset and Middlebury dataset show that the model has better predictive performance for both textureless and repetitive texture regions and has better generalization performance.This method can improve disparity measurement and depth measurement in binocular universal scenes.}, archive = {J_ESWA}, author = {Shengwei Yang and Yuehang Wang and Zenghui Li and Shan Zhou and Zheng Wang and Yining Hu and Lizhe Xie}, doi = {10.1016/j.eswa.2025.129634}, journal = {Expert Systems with Applications}, month = {3}, pages = {129634}, shortjournal = {Expert Syst. Appl.}, title = {Ill-posed regions-aware self-supervised stereo matching with left-right consistency}, volume = {298}, year = {2026}, }
• (2026). TGNet: Texture-enhanced guidance network for RGB-D salient object detection. ESWA, 298, 129633. (www)

  @article{ , abstract = {RGB-D salient object detection achieves salient region localization in complex scenes by fusing RGB images and depth images. Existing methods typically employ two-stream networks to extract features separately followed by cross-modal fusion. However, differences between heterogeneous modalities can easily lead to feature degradation during cross-modal fusion, while the inherent noise interference in low-quality depth maps may generate cumulative effects during multi-stage propagation, severely constraining detection performance. To address these challenges, this paper proposes a texture-enhanced guided network. The core innovations lie in three aspects: during the feature encoding stage, a texture-enhanced module is constructed to utilize high-frequency texture information from RGB images through attention mechanisms for hierarchical optimization of depth features; in the feature fusion stage, a dual-path adaptive interaction module is designed to establish cross-modal semantic correlations via channel-spatial cooperative driving mechanisms, effectively suppressing redundant feature interference; for the decoding reconstruction stage, a dynamic hierarchical guidance mechanism is proposed to drive progressive calibration of low-level spatial details by high-level semantic features through learnable cross-scale transformation modules. Extensive experiments conducted on five benchmark datasets demonstrate that our method achieves competitive performance compared to other approaches.}, archive = {J_ESWA}, author = {Xiaogang Song and Hexiang Huang and Qin Zhao and Xinwei Guo and Xinhong Hei}, doi = {10.1016/j.eswa.2025.129633}, journal = {Expert Systems with Applications}, month = {3}, pages = {129633}, shortjournal = {Expert Syst. Appl.}, title = {TGNet: Texture-enhanced guidance network for RGB-D salient object detection}, volume = {298}, year = {2026}, }
• (2026). Dual-stream temporal-spatial forgery detection via phase-consistent edge features and 3D visual state-space modeling. ESWA, 298, 129632. (www)

  @article{ , abstract = {Deepfake, as a generative technology, has opened up new avenues for the development of the film, television, and art industries. However, its abusive use has triggered serious social security threats, such as infringement of portrait rights and the spread of misinformation, which has drawn widespread attention to research on deepfake detection techniques. Current deep learning-based face forgery detection methods face critical challenges: 1) insufficient focus on common forgery traces leads to poor generalization performance on datasets generated by unknown forgery methods; 2) traditional spatio-temporal feature fusion mechanisms struggle to balance the representational weights of spatial details and temporal dynamics, and exhibit inadequate robustness against post-processing operations like compression and cropping. To address these issues, this paper first designs a phase consistency edge artifact mining module is designed to extract common forgery traces from edge textures by leveraging the deep-phase information of images, significantly enhancing the model’s generalization ability. Second, a multi-frame synthesis strategy is designed to effectively integrate spatial and temporal features while balancing the network’s attention to these two feature domains. Third, a visual state-space model based on 3D scanning is designed, which for the first time employs the Mamba model to analyze spatio-temporal forgery patterns, notably improving the robustness of the model against unknown perturbations. Experimental results on standard benchmarks–FaceForensics++, Celeb-DFv2, WildDeepfake and DFDC(Preview)–demonstrate that the proposed method achieves state-of-the-art performance in three core dimensions: detection accuracy, cross-dataset generalization, and robustness against perturbations.}, archive = {J_ESWA}, author = {Zhong Chen and Siyang Wang and Zuxi Wang}, doi = {10.1016/j.eswa.2025.129632}, journal = {Expert Systems with Applications}, month = {3}, pages = {129632}, shortjournal = {Expert Syst. Appl.}, title = {Dual-stream temporal-spatial forgery detection via phase-consistent edge features and 3D visual state-space modeling}, volume = {298}, year = {2026}, }
• (2026). LSTT: Long short-term feature enhancement transformer for video small object detection. ESWA, 298, 129631. (www)

  @article{ , abstract = {Leveraging temporal information is crucial for small object detection in videos. Existing methods typically incorporate long-term or short-term temporal information uniformly, neglecting distinct cues from different frames that are essential for small object detection. In this paper, we propose LSTT, an end-to-end multi-frame fusion network that concurrently extracts global scene context from long-term frames and fine-grained appearance and motion cues from short-term frames. First, we introduce a progressive spatiotemporal sampling module that sparsely samples long-range frames and densely samples short-range frames. Second, we design a spatiotemporal alignment encoder module to extract frame-level temporal and spatial pixel features. Finally, We propose a long short-term feature aggregation module that employs a dynamic query generator to derive adaptive queries by implicitly modeling motion relationships among short-term frames, and guides a cascaded fusion of aggregated features from long-term, short-term, and current frames to fuse temporal information. Compared to state-of-the-art methods, our LSTT achieves absolute gains of 1.4 % and 2.1 % in detection precision on VisDrone-VID and UAVDT datasets, respectively.}, archive = {J_ESWA}, author = {Jinsheng Xiao and Wenbo Liu and Ruidi Chen and Yuchen Yan and Wei Yang}, doi = {10.1016/j.eswa.2025.129631}, journal = {Expert Systems with Applications}, month = {3}, pages = {129631}, shortjournal = {Expert Syst. Appl.}, title = {LSTT: Long short-term feature enhancement transformer for video small object detection}, volume = {298}, year = {2026}, }
• (2026). Scale-invariant information bottleneck for domain generalization. ESWA, 298, 129628. (www)

  @article{ , abstract = {One significant challenge in deep learning is the inability to effectively generalize to new data whose distribution differs from that of the training data. Hence, domain generalization has received increasing attention in related fields. Classical methods aim to identify an invariant predictor that can recognize invariant representations across all the training domains. However, these methods limit the model to rely solely on invariant representations, which hinders the learning of important finer details. To address this challenge, we propose a Scale-invariant Information Bottleneck (SIB) method to identify both invariant and scale-invariant features. We subsequently introduce a tractable loss function derived from the variational analysis. This novel method captures more detailed information, including fine textures and unique characteristics, while also eliminating irrelevant or spurious representations by using information bottleneck. Finally, extensive experiments conducted on Rotated MNIST, Colored MNIST, Colored Fashion-MNIST, PACS, Office-Home and Camelyon17-WILDS validate the effectiveness of our SIB method in addressing the domain generalization problems. Notably, our approach outperforms 14 existing methods with an average improvement of 4.74 %. More significantly, it surpasses 6 recent, related methods by an average of 2.21 %. Furthermore, we demonstrate the superiority of our method through the analysis of hidden feature maps and representations.}, archive = {J_ESWA}, author = {Mengyao Li and Jiangshe Zhang and Chunxia Zhang and Junmin Liu and Lizhen Ji}, doi = {10.1016/j.eswa.2025.129628}, journal = {Expert Systems with Applications}, month = {3}, pages = {129628}, shortjournal = {Expert Syst. Appl.}, title = {Scale-invariant information bottleneck for domain generalization}, volume = {298}, year = {2026}, }
• (2026). Optimal pricing for transfer of development rights under risk sharing: The context of china’s inter-provincial construction land quota trading. ESWA, 298, 129627. (www)

  @article{ , abstract = {Scientific and rational prices are determinant for the success of transfer of development rights (TDR). Nevertheless, previous studies largely overlook the multifaceted impacts of risks on pricing, hampering market participation and value revelation. This is especially relevant in the context of China’s inter-provincial construction land quota trading due to its broader scope and dynamic complexities. This study addresses this gap by proposing an integrated decision-making framework to identify TDR risk factors and determine the optimal pricing for TDR under risk sharing. Results show that among 28 identified risk factors across the trading lifecycle, pre-transaction (remediation application and remediation acceptance) risk factors exhibit lower weights (0.017 and 0.218) but demand greater responsibility from quota-sending governments (80.5% and 73.0%); quota transfer risk factors hold the highest weight (0.306), with nearly balanced responsibility sharing between trading parties; while post-transaction (remediation acceptance and post monitoring) risk factors (weighted at 0.215 and 0.218) should be borne mainly by quota-receiving governments (64.1% and 60.9%). A paradigmatic trading case study between Muli and Jiashan Counties empirically reveals that risk factors elevate the optimal price to 621171.89 yuan/mu—24.23% above the current national standard price—by increasing costs, reducing profits, decreasing the supply–demand ratio, and complicating ecological compensation. These findings underscore the importance of risk responsibility management and risk-based pricing mechanisms.}, archive = {J_ESWA}, author = {Jia-He Zhou and Yu-Jia Wei and Yu-Ming Zhu and Hong-Li Lin}, doi = {10.1016/j.eswa.2025.129627}, journal = {Expert Systems with Applications}, month = {3}, pages = {129627}, shortjournal = {Expert Syst. Appl.}, title = {Optimal pricing for transfer of development rights under risk sharing: The context of china’s inter-provincial construction land quota trading}, volume = {298}, year = {2026}, }
• (2026). Intelligent urban on-street parking space management for autonomous vehicles. ESWA, 298, 129626. (www)

  @article{ , abstract = {Curbside lanes are valuable spatial assets, with on-street parking, driving, and other travel modes competing for the space. Autonomous vehicle (AV) transport is expected to park at the curbside for diverse purposes, raising conflicts between driving and parking in the city centre. This study presents a framework to determine on-street parking configurations under different traffic flow and parking supply scenarios for the downtown region. The main contribution stems from solving the macro-level parking configuration problem using customised metaheuristics while considering microscopic AV operations. We tested the framework using a road network comprising a downtown central business district and adjacent urban areas. Among the considered metaheuristics, the discrete particle swarm optimisation outperformed the genetic algorithm in minimising network-level travel delays but at the cost of higher computational time. Three main empirical findings are derived. First , parking lanes are more likely to be assigned to edges in downtown areas or those with lower traffic and driving speeds. Second , high parking supply negatively affects the macroscopic fundamental diagram by increasing congestion and reducing flow efficiency, but such an effect diminishes in congested networks. Third , there exists an optimal parking supply level (40 % in the case study) for most flow rate conditions that can help reduce congestion. The proposed framework was validated through a case study in Midtown Manhattan, New York City. This study provides valuable insights for urban and transportation agencies to manage on-street parking lane assignments to balance parking and driving demands in the AV transport era. The approach has broader applicability as it is transferable to human-driven vehicles and mixed autonomy scenarios.}, archive = {J_ESWA}, author = {Qiming Ye and Prateek Bansal and Yuxiang Feng and Simon Hu and Panagiotis Angeloudis}, doi = {10.1016/j.eswa.2025.129626}, journal = {Expert Systems with Applications}, month = {3}, pages = {129626}, shortjournal = {Expert Syst. Appl.}, title = {Intelligent urban on-street parking space management for autonomous vehicles}, volume = {298}, year = {2026}, }
• (2026). Cross-session interest extraction for recommendation. ESWA, 298, 129625. (www)

  @article{ , abstract = {In recommender systems, due to device privacy restrictions, sometimes we can only obtain anonymous users’ interaction behavior within a single session. This type of recommendation is called session-based recommendation. Modeling users’ interest based on session data is one of the core issues in session-based recommendation. However, most existing methods only model users’ interest within individual sessions, neglecting information propagation across sessions. This paper addresses this challenge by designing a contrastive learning module based on clustering to model inter-session information propagation. Specifically, in addition to propagating information within sessions using hypergraph convolution, a cluster algorithm is applied to group all nodes across sessions. Then a contrastive learning loss is designed based on the clustering results to facilitate information propagation across sessions, thereby explicitly modeling the semantic similarity of similar items across different sessions. We call our model Clustering Hypergraph Neural Network (CluHNN). CluHNN explicitly learns the correlation between similar items across different sessions, improving the quality of item representations and, consequently, yielding better interest representations through cross-session information propagation. Experimental results on two real-world datasets show the effectiveness of the proposed CluHNN. For example, in terms of MRR@20, CluHNN achieves significant improvements of 1.3 % and 2.0 % relative gains over the strongest baseline, respectively.}, archive = {J_ESWA}, author = {Jin Jin and Chaoqun Li and Liangxiao Jiang}, doi = {10.1016/j.eswa.2025.129625}, journal = {Expert Systems with Applications}, month = {3}, pages = {129625}, shortjournal = {Expert Syst. Appl.}, title = {Cross-session interest extraction for recommendation}, volume = {298}, year = {2026}, }
• (2026). Incorporating estimated depth maps and multi-modal pretraining to improve salient object detection in optical remote sensing images. ESWA, 298, 129624. (www)

  @article{ , abstract = {As a burgeoning theme in optical remote sensing image (ORSI) analysis, salient object detection (SOD) plays a vital role in traffic monitoring, agriculture, disaster management, and other fields. However, the existing ORSI-SOD methods are all single-modal (RGB images primarily), which suffer from performance drop when facing complex scenes (e.g., intricate backgrounds, low contrast scenes, and similar objects). To address this challenge, we introduce estimated depth map to complement RGB image in ORSI-SOD for the first time, which provides 3D geometric cues to improve detection accuracy in complex scenes, thus advancing ORSI-SOD from single-modal to multi-modal. Furthermore, we design a novel pretraining framework: multi-modal reconstructed image pretraining (MMRIP) to pretrain SOD model in multi-modal ORSI-SOD. MMRIP initially utilizes a masked autoencoder (MAE) to restore the masked RGB image; subsequently, it feeds the restored RGB image and clean depth map to the SOD model to generate the saliency map, which can help SOD model more effectively integrate cross modal information and extract better feature. Besides, we present a simple RGB-D SOD model, namely SimSOD, which is pretrained by MMRIP for ORSI-SOD. SimSOD has two major components: DFormer (encoder) and MLP head (decoder). Specifically, we first input RGB image and depth data into the encoder to generate four multi-scale features, then use the decoder to fuse these features and yield the prediction result. Without bells and whistles, our proposed method outperforms the state-of-the-art methods on three public ORSI-SOD datasets. The code can be accessed at: https://github.com/Voruarn/MMRIP .}, archive = {J_ESWA}, author = {Yuxiang Fu and Wei Fang}, doi = {10.1016/j.eswa.2025.129624}, journal = {Expert Systems with Applications}, month = {3}, pages = {129624}, shortjournal = {Expert Syst. Appl.}, title = {Incorporating estimated depth maps and multi-modal pretraining to improve salient object detection in optical remote sensing images}, volume = {298}, year = {2026}, }
• (2026). An effective new penetration test approach to detect web attacks on web applications. ESWA, 298, 129623. (www)

  @article{ , abstract = {As web applications increasingly involve sensitive transactions such as e-commerce, online banking, and public services, they have become primary targets for cyberattacks. Therefore, web application penetration testing is vital for discovering and protecting against vulnerabilities in web-based systems. This study introduces an automated penetration testing tool that systematically applies comprehensive penetration testing methodologies to effectively identify and mitigate web application vulnerabilities. The proposed tool uses a hybrid approach that includes automated and manual testing phases for multiple attacks, including SQL Injection, Cross-Site Scripting, and Cross-Site Request Forgery. System diversity is increased, and the penetration testing process is enriched using Python scripts and APIs. The tool provides an effective mechanism for uncovering critical vulnerabilities by simulating real-world attacker behavior. Practical evaluations on various web applications demonstrate the tool’s ability to identify vulnerabilities and increase system resilience. This research will help developers and security engineers apply this automated, specialized approach to security as our digital environment becomes more connected.}, archive = {J_ESWA}, author = {Muhammed Onur Kaya and Huseyin Alperen Dagdogen and Mehmet Ozdem and Resul Das}, doi = {10.1016/j.eswa.2025.129623}, journal = {Expert Systems with Applications}, month = {3}, pages = {129623}, shortjournal = {Expert Syst. Appl.}, title = {An effective new penetration test approach to detect web attacks on web applications}, volume = {298}, year = {2026}, }
• (2026). An approach for linking dynamic network information models based on ontology matching. ESWA, 298, 129622. (www)

  @article{ , abstract = {Dynamic network information models are typically heterogeneous and isolated systems that impede effective interoperability, significantly hindering end-to-end service integration and data sharing across network segments. To address this challenge, we propose a new approach for linking heterogeneous dynamic network models based on ontology matching, which can be applied in various domains utilizing dynamic networks. For ontologies matching we use different existing duplicate detection algorithms but we reduce the computational complexity of ontology matching due to splitting initial set of matched entities into a number of subsets using domain knowledge. Using telecommunications as case study, we represent operator networks as knowledge graphs and match them with standardized model ontologies using business process context to create an Extended Operator Network Ontology. Our approach ensures linking of dynamic network models used in operators information systems that is of primary importance for implementing complex business processes, and providing integrated services while maintaining existing models.}, archive = {J_ESWA}, author = {Tianxing Man and Igor Kulikov and Jiafeng Yang and Nataly Zhukova}, doi = {10.1016/j.eswa.2025.129622}, journal = {Expert Systems with Applications}, month = {3}, pages = {129622}, shortjournal = {Expert Syst. Appl.}, title = {An approach for linking dynamic network information models based on ontology matching}, volume = {298}, year = {2026}, }
• (2026). An improved differential evolution algorithm combined with vector NFP and mixed-integer programming for solving 2D irregular layout problem. ESWA, 298, 129621. (www)

  @article{ , abstract = {The two-dimensional irregular layout problem, which involves placing convex or non-convex components within a confined boundary without overlaps, is NP-complete and widely encountered in industrial applications such as glass cutting, garment manufacturing, and packaging. To overcome the limitations of existing methods in computational efficiency and material utilization, we propose a new hybrid algorithm IDE-V-NFP-MIP: (1) An improved differential evolution (IDE) algorithm combines the memory mechanism to guide the crossover and mutation operations; (2) A vector No-Fit Polygon (V-NFP) algorithm effectively handles complex geometric constraints, including voids and degradation; (3) A mixed-integer programming (MIP) model ensures accurate layout and non-overlapping constraints. Experimental results demonstrate superior performance: IDE ranked first in CEC2022 Friedman tests, while practical applications show 22.10% reduction in board length and 41.47% improvement in filling rate. The framework successfully handles real-world garment cutting applications and large-scale problems up to 1,280 polygons, demonstrating significant improvements in both computational efficiency and solution quality for industrial layout optimization. The source code for the algorithm is available at https://github.com/xhj-6/IDE-V-NFP-MIP .}, archive = {J_ESWA}, author = {Huijie Xu and Qifang Luo and Yongquan Zhou}, doi = {10.1016/j.eswa.2025.129621}, journal = {Expert Systems with Applications}, month = {3}, pages = {129621}, shortjournal = {Expert Syst. Appl.}, title = {An improved differential evolution algorithm combined with vector NFP and mixed-integer programming for solving 2D irregular layout problem}, volume = {298}, year = {2026}, }
• (2026). Intelligent localization of FDIA in smart grids: A multi-level wavelet spatio-temporal graph embedding approach. ESWA, 298, 129620. (www)

  @article{ , abstract = {In the realm of smart grid security, the precise identification of False Data Injection Attack (FDIA) is crucial for ensuring the stable operation of power systems. Existing approaches for handling measurement data often overlook the correlation between local time–frequency variations caused by FDIA nodes and global spatial information in analyzing measurement data, leading to inaccurate localization. To address this issue, we propose a novel approach: a multilevel wavelet spatio-temporal map embedded FDIA localization method. Initially, a multi-resolution time–frequency signal decomposition model is utilized to separate the time–frequency mutation signals induced by FDIA from the measurement data using fast wavelet transform. Subsequently, a multi-channel time–frequency feature extraction technique is developed to capture the mutation characteristics of FDIA in time–frequency signals. This involves extracting detailed features of the time–frequency signals pre and post-attack via a multi-channel convolution operation encompassing “temporal-local-global” aspects. Finally, we propose an FDIA localization model based on multi-level graph wavelet embedding. The model embeds spatio-temporal information into time–frequency features via graph wavelet convolution and builds a spatio-temporal dependency map through multi-level neighborhood sampling. To mitigate measurement loss and noise, graph smoothing regularization and graph dropout are introduced during training. A graph attention mechanism further captures spatio-temporal dependencies among nodes, enabling accurate FDIA localization. Experimental results verify the effectiveness of the proposed method.}, archive = {J_ESWA}, author = {Zhaoyang Qu and Feng Liang and Nan Qu and Tao Jiang and Xiaoyu Xu}, doi = {10.1016/j.eswa.2025.129620}, journal = {Expert Systems with Applications}, month = {3}, pages = {129620}, shortjournal = {Expert Syst. Appl.}, title = {Intelligent localization of FDIA in smart grids: A multi-level wavelet spatio-temporal graph embedding approach}, volume = {298}, year = {2026}, }
• (2026). AI-driven 5G-IoT optimization: Q-learning for real-time energy and network resource management. ESWA, 298, 129619. (www)

  @article{ , abstract = {In recent times, the integration of the 5G-enabled Internet of Things has revolutionized through high-speed data transmission, ultra-low latency, and interconnectivity of massive devices. However, the proliferation of 5G-enabled Internet of Things introduces major challenges, such as energy inefficiency and unreliable data delivery in the resource constrained Internet of Things devices. This research proposes a novel Q-Learning-based optimization framework tailored to address these challenges by integrating Radio Frequency energy harvesting, adaptive beamforming, and dynamic resource allocation within the massive Multiple-Input-Multiple-Output system. The proposed model utilizes reinforcement learning to manage the network resources including modulation schemes, beamforming, and energy allocation. By modeling the optimization problem as a Markov Decision Process, the proposed framework dynamically adapts to real-time network conditions to enhance energy efficiency, reliable data delivery, and throughput. The experimental validation demonstrates that the Q-Learning-based strategy effectively optimizes the energy efficiency as well as data transmission and achieves a higher energy efficiency of 98.87 %, higher packet delivery ratio of 98.85 %, lower latency of 1.5 ms, and higher throughput of 200Mbps compared to existing methodologies. This result indicates that the proposed Q-Learning-based framework has the potential to enhance the sustainability and reliability of the 5G-enabled Internet of Things.}, archive = {J_ESWA}, author = {Bavethra Murthy and Palani Uthirapathy}, doi = {10.1016/j.eswa.2025.129619}, journal = {Expert Systems with Applications}, month = {3}, pages = {129619}, shortjournal = {Expert Syst. Appl.}, title = {AI-driven 5G-IoT optimization: Q-learning for real-time energy and network resource management}, volume = {298}, year = {2026}, }
• (2026). Multi-behavioral recommendation algorithm based on decoupled graph convolution. ESWA, 298, 129618. (www)

  @article{ , abstract = {Traditional recommendation models primarily rely on display feedback and typically utilize a single type of user-item interaction data, which often results in significant data sparsity issues. In contrast, multi-behavioral recommendation models leverage various behaviors such as browsing, favoriting, and other interactions. These additional behaviors help improve the prediction of user-item interactions. Existing multi-behavioral recommendation methods often overlook the potential factors influencing multi-behavioral interactions and the differences between various behavior types. In this study, we introduce a multi-behavioral recommendation algorithm utilizing decoupled graph convolution (MBR-DGC), which effectively mitigates the data sparsity of the target behaviors and improves recommender system performance by capturing the differences between the semantics of different behaviors. Specifically, we construct multiple non-overlapping independent isomorphic graphs and separate potential factors affecting the interactions among users, items, and behaviors using decoupled convolutional networks to reconstruct the node features of users in different behaviors. Afterwards, multi-behavioral features of users are aggregated using contrastive learning to achieve personalized multi-behavioral information aggregation. Experimental results on multiple datasets show that MBR-DGC effectively leverages multi-behavioral data, significantly enhancing recommendation performance compared to other state-of-the-art methods.}, archive = {J_ESWA}, author = {Xu Yu and Pengju Ding and Jie Yu and Junyu Lin and Lei Guo and Guanfeng Liu and Liang Xi}, doi = {10.1016/j.eswa.2025.129618}, journal = {Expert Systems with Applications}, month = {3}, pages = {129618}, shortjournal = {Expert Syst. Appl.}, title = {Multi-behavioral recommendation algorithm based on decoupled graph convolution}, volume = {298}, year = {2026}, }
• (2026). An information aggregation decision making method for UAV swarm intelligence system based on joint communication and proximal strategy. ESWA, 298, 129617. (www)

  @article{ , abstract = {The swarm intelligence aggregation system represents a key capability in current-generation UAV swarm, demonstrating robust collective intelligence. Currently, leveraging Multi-Agent Deep Reinforcement Learning (MADRL) offers a promising approach for building UAV swarm intelligence aggregation systems. However, the MADRL methods are difficult to cope with the challenge of exponential increase in computation when facing the collaboration problem of large-scale swarms, and the agents also have the problem of partial observability of the environment. This paper proposes an Information Aggregation Decision Method for UAV swarm based on Joint Communication and Proximal Strategy (IADM-JCPS). This method designs a communication information aggregation (CIA) network to enable UAVs to gather observation information from neighbor UAVs, and uses the attention mechanism to screen important information. Then, the aggregated information is used as part of the input of the policy network to increase the information diversity of the decision-making process. Finally, the gradient clipping mechanism is used to trim the policy gradient to enhance the stability of the training process. A UAV swarm multi-target tracking (MTT) mission scenario is designed to verify the effectiveness of the proposed IADM-JCPS algorithm. Experimental results show that the proposed algorithm is superior to the baseline algorithm in terms of task collaboration and scalability.}, archive = {J_ESWA}, author = {Zhaotian Wei and Ruixuan Wei}, doi = {10.1016/j.eswa.2025.129617}, journal = {Expert Systems with Applications}, month = {3}, pages = {129617}, shortjournal = {Expert Syst. Appl.}, title = {An information aggregation decision making method for UAV swarm intelligence system based on joint communication and proximal strategy}, volume = {298}, year = {2026}, }
• (2026). Learning-based trajectory planning for AGVs in dynamic environment. ESWA, 298, 129616. (www)

  @article{ , abstract = {In this work, we present a learning-based framework for rapid trajectory planning of autonomous ground vehicles (AGVs) in dynamic environments. The approach integrates optimization techniques with deep learning to design a real-time planner capable of generating kinematically feasible trajectories. A continuous iterative method is first developed for dataset construction, enabling efficient generation of optimal trajectory sets. Based on this dataset, a neural network is trained to learn the mapping between AGV states and actions while capturing their temporal dependencies. During online planning, the trained model produces decision actions from the current state and sensor feedback, enabling real-time planning of safe and feasible trajectories. Results demonstrate the effectiveness of the proposed framework.}, archive = {J_ESWA}, author = {Runda Zhang and Zhida Xing and Senchun Chai and Yuanqing Xia and Runqi Chai}, doi = {10.1016/j.eswa.2025.129616}, journal = {Expert Systems with Applications}, month = {3}, pages = {129616}, shortjournal = {Expert Syst. Appl.}, title = {Learning-based trajectory planning for AGVs in dynamic environment}, volume = {298}, year = {2026}, }
• (2026). Model-agnostic post-hoc explainability for recommender systems. ESWA, 298, 129608. (www)

  @article{ , abstract = {Recommender systems often benefit from complex feature embeddings and deep learning algorithms, which deliver sophisticated recommendations that enhance user experience, engagement, and revenue. However, these methods frequently reduce the interpretability and transparency of the system. In this research, we develop a systematic application, adaptation, and evaluation of deletion diagnostics in the recommender setting. The method compares the performance of a model to that of a similar model trained without a specific user or item, allowing us to quantify how that observation influences the recommender, either positively or negatively. To demonstrate its model-agnostic nature, the proposal is applied to both Neural Collaborative Filtering (NCF), a widely used deep learning-based recommender, and Singular Value Decomposition (SVD), a classical collaborative filtering technique. Experiments on the MovieLens and Amazon Reviews datasets provide insights into model behavior and highlight the generality of the approach across different recommendation paradigms.}, archive = {J_ESWA}, author = {Irina Arévalo and Jose L. Salmeron}, doi = {10.1016/j.eswa.2025.129608}, journal = {Expert Systems with Applications}, month = {3}, pages = {129608}, shortjournal = {Expert Syst. Appl.}, title = {Model-agnostic post-hoc explainability for recommender systems}, volume = {298}, year = {2026}, }
• (2026). AERO-net: AutoEncoder-driven residual optimization network for latent bias correction of WRF-ROMS. ESWA, 298, 129607. (www)

  @article{ , abstract = {Reliable assessment of precipitation is crucial for incorporating meteorological and hydrological research into industrial and agricultural applications. Accurately estimating precipitation is a challenging task. In addressing this problem, we propose to develop AERO-Net, a novel deep learning framework designed to correct spatial, temporal, and amplitude biases in WRF-ROMS precipitation data. The integration of the Weather Research and Forecasting (WRF) model with the Regional Ocean Modeling System (ROMS) makes it a valuable tool for precipitation forecasting. AERO-Net incorporates autoencoders (AEs) for handling fluctuation and generalizing latent space representations, a latent module (LM) for transforming WRF-ROMS data into bias-corrected representations, a residual module (RM) for error minimization via boost, and a calibration module (CM) for improving near-zero precipitation. Empirical results show that AERO-Net achieves a balanced error reduction across precipitation cohorts grouped by intensity, reducing the macro-averaged root mean square error (macro RMSE) by 3.6 mm/day and the macro-averaged mean absolute deviation (macro MAD) by 0.68 mm/day compared to the original WRF-ROMS. AERO-Net is seen to improve the correlation coefficient (CC) by 26.32 %, increasing it from 0.38 to 0.48, in comparison to the original WRF-ROMS. These findings underscore its potential as an effective solution for enhancing precipitation estimates in high-resolution modeling systems.}, archive = {J_ESWA}, author = {Passin Pornvoraphat and Kanoksri Sarinnapakorn and Ken-Ichi Fukui and Peerapon Vateekul}, doi = {10.1016/j.eswa.2025.129607}, journal = {Expert Systems with Applications}, month = {3}, pages = {129607}, shortjournal = {Expert Syst. Appl.}, title = {AERO-net: AutoEncoder-driven residual optimization network for latent bias correction of WRF-ROMS}, volume = {298}, year = {2026}, }
• (2026). Contrastive learning and physics oriented evaluation for advanced segmentation in electron tomography. ESWA, 298, 129606. (www)

  @article{ , abstract = {Deep learning methods are now achieving strong results for segmentation tasks, and the standard metric for evaluating methods is the Intersection over Union (IOU). However, we show in this paper that IOU is not efficient in evaluating the quality of segmentation for electron tomography (ET) images of zeolites. We perform a physics-oriented evaluation to ensure that the segmentation results yield coherent physical measures. We also formalize Mixed Supervised / Self-Supervised Contrastive Learning Segmentation (M3S-CLS), a semi-supervised approach using a contrastive learning approach that uses expert annotations to train the neural network model. A detailed comparison of this method with a standard cross-entropy-based model is provided. In addition, we publish a database of five fully segmented ET volumes along with corresponding baseline results. The code and the database is available at http://gitlab.univ-st-etienne.fr/labhc-iscv/M3S-CLS .}, archive = {J_ESWA}, author = {Cyril Li and Christophe Ducottet and Maxime Moreaud and Sylvain Desroziers and Valentina Girelli Consolaro and Virgile Rouchon and Ovidiu Ersen}, doi = {10.1016/j.eswa.2025.129606}, journal = {Expert Systems with Applications}, month = {3}, pages = {129606}, shortjournal = {Expert Syst. Appl.}, title = {Contrastive learning and physics oriented evaluation for advanced segmentation in electron tomography}, volume = {298}, year = {2026}, }
• (2026). DSANet: Deep surrounding-aware network for camouflaged object detection via cross-refinement mirror strategy. ESWA, 298, 129605. (www)

  @article{ , abstract = {Camouflaged objects often closely resemble their surroundings, causing standard RGB images to be confounded by background, texture, and color variations. This often leads to incomplete or absent target segmentation, reducing overall accuracy. To address this issue, we present a Deep Surrounding-Awareness Mirror Network (DSANet) for camouflaged object detection, leveraging depth information to expose objects incongruent with their environment, thus improving localization accuracy. First, a Convolutional Spatial Gating module processes batched RGB and depth inputs, suppressing extraneous background noise while isolating fine-grained segmentation and structural features and unifying channel representation. Subsequently, a Deep Surrounding-Awareness Localization module and a Contour-Guided Integrity Aggregation module collaboratively refine and merge multi-level features, focusing on the global form of camouflaged objects while iteratively enhancing segmentation detail. Finally, a Guided Residual Channel Attention module further refines low-layer structural cues. Extensive experiments on ten challenging benchmark datasets using four widely used evaluation metrics demonstrate that our method exhibited superior performance, outperforming 40 state-of-the-art methods. The results demonstrate the versatility of our model. The source code and results of our method are available at https://github.com/lixu11/DSANet.}, archive = {J_ESWA}, author = {Xu Li and Xiaosheng Yu and Peng Chen}, doi = {10.1016/j.eswa.2025.129605}, journal = {Expert Systems with Applications}, month = {3}, pages = {129605}, shortjournal = {Expert Syst. Appl.}, title = {DSANet: Deep surrounding-aware network for camouflaged object detection via cross-refinement mirror strategy}, volume = {298}, year = {2026}, }
• (2026). FTUAttack: Feature truncation unrestricted attack based on stable diffusion model. ESWA, 298, 129604. (www)

  @article{ , abstract = {In the field of adversarial example generation and defense, compared to restricted attacks with L p -norm constraints, unrestricted attacks without L p -norm constraints emanate better visual imperceptibility. Existing unrestricted attacks typically manipulate the semantic content of examples (e.g. texture or color) to generate adversarial examples. However, current works usually ignore multifaceted features or loss optimization strategy, which limits attack performance. In this paper, we draw inspiration from stable diffusion model and propose a unrestricted attack method called Feature Truncation Unrestricted Attack (FTUAttack) to achieve both better transferability and imperceptibility. Specifically, we promote the performance of unrestricted attacks from the perspectives of both diffusion principle and feature truncation for the first time. Firstly, we propose a Global Deep Feature Extractor (GDFE) module to truncate global feature for the subsequent diffusion denoising process. Secondly, to further boost the transferability, we design a novel Critical Latent Feature Extractor (CLFE) module to obtain critical local feature that need to be truncated during the denoising process and investigate the influence of the different segmentation ways on critical local feature. Thirdly, we propose Multi-Loss Fusion (MLF) strategy to balance the conflict between perturbations and examples’ quality by guiding the optimization direction. Extensive experiments on various model structures and datasets demonstrate the superiority of our attack over the existing attack methods.}, archive = {J_ESWA}, author = {Shaojie Han and Gangzheng Zhai and Kun Chen and Shihui Zhang}, doi = {10.1016/j.eswa.2025.129604}, journal = {Expert Systems with Applications}, month = {3}, pages = {129604}, shortjournal = {Expert Syst. Appl.}, title = {FTUAttack: Feature truncation unrestricted attack based on stable diffusion model}, volume = {298}, year = {2026}, }
• (2026). Spatial-temporal hierarchical decoupled masked autoencoder: A self-supervised learning framework for electrocardiogram. ESWA, 298, 129603. (www)

  @article{ , abstract = {The difficulty of labeling Electrocardiogram (ECG) has prompted researchers to use self-supervised learning to enhance diagnostic performance. Masked autoencoders (MAE) are a mainstream paradigm where models learn a latent representation of the signal by reconstructing masked portions of the ECG. However, existing methods lack a specific design for the spatial–temporal characteristics of ECG. Specifically, leads represent spatial projections of cardiac activity, while timestamps capture temporal patterns, and the two correspond to different axes of information. Existing MAE frameworks tend to unify them prematurely, potentially weakening critical local dependencies. In this paper, we propose a Spatial-Temporal Hierarchical Decoupled Masked Autoencoder (STHD-MAE). This framework decouples ECG into isolated leads or time steps in the shallow layer to capture local dependencies with different views, then aligns spatial–temporal representations and re-establishes global dependencies in the deep layer to comprehensively represent pathological information. We also design a medical report fusion module during pre-training, which uses cross-attention to align the ECG report text encoded by a medical language model with the signal’s latent representation, thereby guiding the encoder to focus on pathological information through implicit cross-modal learning. We validate the effectiveness of STHD-MAE on multiple downstream classification and reconstruction tasks. The results show that STHD-MAE outperforms existing self-supervised learning methods by approximately 2% in F1-scores for both coarse-grained and fine-grained classification performance, and its reconstruction quality also exceeds the baseline generative model.}, archive = {J_ESWA}, author = {Xiaoyang Wei and Zhiyuan Li and Yuanyuan Tian and Mengxiao Wang and Yanrui Jin and Weiping Ding and Chengliang Liu}, doi = {10.1016/j.eswa.2025.129603}, journal = {Expert Systems with Applications}, month = {3}, pages = {129603}, shortjournal = {Expert Syst. Appl.}, title = {Spatial-temporal hierarchical decoupled masked autoencoder: A self-supervised learning framework for electrocardiogram}, volume = {298}, year = {2026}, }
• (2026). GARE-net: Geometric contextual aggregation and regional contextual enhancement network for image-text matching. ESWA, 298, 129602. (www)

  @article{ , abstract = {The local correspondence learning has gained increasing attention in image-text matching, which establishes fine-grained alignments between image regions and textual words to improve both interpretability and accuracy. While these approaches have made significant progress in identifying meaningful semantic correspondences, one critical limitation persists in current methods, i.e., overlooking the crucial spatial position information of visual regions in cross-modal interaction. To address this challenge, we propose a novel Geometric contextual Aggregation and Regional contextual Enhancement Network (GARE-Net) that introduces two innovative components: the Geometric Contextual Feature Aggregation (GCFA) module and the Regional Contextual Feature Enhancement (RCFE) module. Specifically, GCFA generates the spatial geometric information of visual regions to enhance the region features by feature aggregation. RCFE further refines the aggregated region features by constructing a region graph and graph convolution. Extensive experiments and analyses are conducted on Flickr30k and MSCOCO to evaluate the importance of our framework. The results demonstrate the superiority of our method in image-text matching. Moreover, the ablation studies and visualization case studies also highlight the importance of geometric contextual feature aggregation and regional contextual feature enhancement. The code is available at https://github.com/chinaBoy123/GARE-Net .}, archive = {J_ESWA}, author = {Fangming Zhong and Tao Zhou and Zhikui Chen and Suhua Zhang}, doi = {10.1016/j.eswa.2025.129602}, journal = {Expert Systems with Applications}, month = {3}, pages = {129602}, shortjournal = {Expert Syst. Appl.}, title = {GARE-net: Geometric contextual aggregation and regional contextual enhancement network for image-text matching}, volume = {298}, year = {2026}, }
• (2026). GSFF-SLAM: 3D semantic gaussian splatting SLAM via feature field. ESWA, 298, 129601. (www)

  @article{ , abstract = {Semantic-aware 3D scene reconstruction is essential for autonomous robots to perform complex interactions. Semantic SLAM, an online approach, integrates pose tracking, geometric reconstruction, and semantic mapping into a unified framework, shows significant potential. However, existing systems, which rely on 2D ground truth priors for supervision, are often limited by the sparsity and noise of these signals in real-world environments. To address this challenge, we propose GSFF-SLAM, a novel dense semantic SLAM system based on 3D Gaussian Splatting that leverages feature fields to achieve joint rendering of appearance, geometry, and N-dimensional semantic features. By independently optimizing feature gradients, our method supports semantic reconstruction using various forms of 2D priors, particularly sparse and noisy signals. Experimental results demonstrate that our approach outperforms previous methods in both tracking accuracy and photorealistic rendering quality. When utilizing 2D ground truth priors, GSFF-SLAM achieves state-of-the-art semantic segmentation performance with 95.03 % mIoU.}, archive = {J_ESWA}, author = {Zuxing Lu and Xin Yuan and Shaowen Yang and Jingyu Liu and Changyin Sun}, doi = {10.1016/j.eswa.2025.129601}, journal = {Expert Systems with Applications}, month = {3}, pages = {129601}, shortjournal = {Expert Syst. Appl.}, title = {GSFF-SLAM: 3D semantic gaussian splatting SLAM via feature field}, volume = {298}, year = {2026}, }
• (2026). Quantum probabilistic group consensus decision-making model based on matrix fluctuation grey correlation for engineering bid evaluation. ESWA, 298, 129600. (www)

  @article{ , abstract = {The existing quantum group decision-making models face significant challenges in the bid evaluation of engineering projects, including the strong subjectivity of expert evaluations, the difficulty in aggregating expert opinions, the large gap of expert opinions, and the complexity of expert psychological behaviors. To address these issues, this paper proposes a novel quantum probabilistic group consensus decision-making model based on matrix fluctuation grey correlation. Firstly, a quantum Bayesian network is constructed to aggregate expert opinions and capture the interference effect among experts. Secondly, the matrix fluctuation grey correlation degree is defined and applied to the calculation of quantum interaction terms that reflect the intricate psychological behavior of experts. Subsequently, a decision item search model is proposed and applied to adjust preferences during the consensus reaching process, thereby narrowing the gap of expert opinions. The consensus effect optimization model is utilized to determine optimal values for unknown parameters within this process, effectively reducing the subjectivity of expert evaluations. Finally, the proposed model is applied to a bid evaluation of bridge anti-collision engineering project, which verifies the feasibility and effectiveness of the model, and evaluates the stability and superiority of the model through sensitivity analysis and comparative analysis.}, archive = {J_ESWA}, author = {Jiuru Zhu and Xinping Xiao and Congjun Rao}, doi = {10.1016/j.eswa.2025.129600}, journal = {Expert Systems with Applications}, month = {3}, pages = {129600}, shortjournal = {Expert Syst. Appl.}, title = {Quantum probabilistic group consensus decision-making model based on matrix fluctuation grey correlation for engineering bid evaluation}, volume = {298}, year = {2026}, }
• (2026). DPCND: A dual-population based evolutionary approach for critical node detection problem in complex networks. ESWA, 298, 129599. (www)

  @article{ , abstract = {Critical node detection is an important tool for measuring network robustness. The main purpose of critical node detection is to detect a set of nodes that cause the greatest damage to the network connectivity, and it has been applied in many fields such as social network analysis and traffic network management. As a classic non-deterministic polynomial time complete problem, critical node detection faces enormous challenges with the continuous expansion of network size. The existing methods are difficult to achieve a good balance between effectiveness and efficiency, especially when the scale of complex networks becomes larger. To this end, this paper proposes a dual population based critical node detection method (DPCND) to effectively and efficiently obtain a set of critical nodes, which utilizes the co-evolution of auxiliary population generated from reduced graph and main population generated from original graph to find the optimal solution. In the proposed algorithm, a dual population interaction mechanism consists of influence and expansion strategies is proposed for information exchange, where the influence strategy transfers candidate good solutions from the auxiliary population to the main population to improve search efficiency, and the expansion strategy provides node information of the main population to guide the expansion of search space for the auxiliary population. Finally, the experimental results on 20 real-world complex networks clearly demonstrate the effectiveness of the proposed algorithm comparing to the state-of-the-arts.}, archive = {J_ESWA}, author = {Lei Zhang and Xinyi Feng and Yuanyuan Ge and Zhanpeng Wang and Haipeng Yang}, doi = {10.1016/j.eswa.2025.129599}, journal = {Expert Systems with Applications}, month = {3}, pages = {129599}, shortjournal = {Expert Syst. Appl.}, title = {DPCND: A dual-population based evolutionary approach for critical node detection problem in complex networks}, volume = {298}, year = {2026}, }
• (2026). A two-stage evolutionary algorithm with restart scheme for an integrated robot-task-scheduling and vehicle-dispatch-scheduling problem. ESWA, 298, 129598. (www)

  @article{ , abstract = {In modern agriculture, efficiently picking and distributing fresh product is crucial for the competitiveness of smart farms within the globalized agricultural market. However, the integrated scheduling problem involving both picking and distribution processes has received limited attention in existing research. To bridge this gap, this study establishes a mathematical model with dual objectives: (1) minimizing the picking completion time and (2) reducing penalties incurred due to early or delayed deliveries. A novel two-stage evolutionary algorithm incorporating a restart mechanism is proposed to effectively balance the optimization of these objectives with a high degree of consistency. The algorithm features an efficient encoding scheme and advanced genetic operators, specifically designed to enhance exploration and exploitation based on the characteristics of the problem. A comprehensive set of test instances is generated and the proposed method is benchmarked against several state-of-the-art metaheuristics from the literature. Experimental results demonstrate that the proposed algorithm outperforms the competing approaches by a significant margin for solving the problem under consideration.}, archive = {J_ESWA}, author = {Yiran Pan and Xuan He and Nan Li and Zhonghua Miao}, doi = {10.1016/j.eswa.2025.129598}, journal = {Expert Systems with Applications}, month = {3}, pages = {129598}, shortjournal = {Expert Syst. Appl.}, title = {A two-stage evolutionary algorithm with restart scheme for an integrated robot-task-scheduling and vehicle-dispatch-scheduling problem}, volume = {298}, year = {2026}, }
• (2026). An initial value insensitive method for phase equilibrium calculation: Constrained quadratic interpolation optimization algorithm. ESWA, 298, 129597. (www)

  @article{ , abstract = {The successful simulation of constrained differential evolution (CDE) algorithm for solving phase equilibrium calculation has first verified that heuristic optimization algorithms are effective ways to solve this kind of problems. Their insensitivity to initial values overcomes the limitations associated with two kinds of traditional methods, i.e., direct solution methods based on Newton’s method and indirect solution methods based on thermodynamic principles. This article proposes a constrained quadratic interpolation optimization algorithm (CQIO) for obtaining the satisfactory solutions of phase equilibrium calculation under given volume, temperature, and moles (NVT-flash). The proposed CQIO regards the total Helmholtz free energy of a NVT-flash problem as its objective function, while the moles vector and volume of a certain phase as its decision variables. The consistency between the four cases’ experimental results of CQIO and those of published articles demonstrates the effectiveness of CQIO in solving NVT-flash problems. Then the computational overhead and algorithmic stability of CQIO were analyzed. In Cases 1, 2 and 3, the average CPU time of CQIO compared to CDE has increased by 46.98 % , 54.56 % and 21.02 % respectively. The Std values of CQIO are significantly smaller than those of CDE in all cases except for Example 2. The proposed CQIO greatly promotes the application of heuristic algorithm in the field of phase equilibrium calculation.}, archive = {J_ESWA}, author = {Wangyu Tong and Baoduo Su and Yaqian Zhan}, doi = {10.1016/j.eswa.2025.129597}, journal = {Expert Systems with Applications}, month = {3}, pages = {129597}, shortjournal = {Expert Syst. Appl.}, title = {An initial value insensitive method for phase equilibrium calculation: Constrained quadratic interpolation optimization algorithm}, volume = {298}, year = {2026}, }
• (2026). Knowledge graph life cycle for cognitive agents – A case study on automated negotiation in smart grids. ESWA, 298, 129596. (www)

  @article{ , abstract = {Knowledge Graphs (KGs) can enhance cognitive artificial agents by improving their semantic understanding, adaptability to dynamic contexts, explainability, continuous learning, and informed, case-based decision-making. However, the development of any KG rarely follows an explicit and structured procedure to improve its consistency, modifiability, and interoperability. This ultimately restricts the utility of KGs for real-world agentic AI systems. To overcome this limitation, this paper proposes a life cycle for developing and evolving KGs in domain-specific applications, encompassing three phases to i) conceptualize the problem domain and competency questions, ii) formalize a schema using ontologies, and iii) implement the KG to foster agents’ cognition through queries and graph data science. The proposed approach is validated with a case study on context-aware automated negotiations within smart grids, where agents negotiate for energy trading while considering private and contextual circumstances. Agents may take actions with or without the aid of a KG, and can adopt one of three negotiation strategy configurations: heuristic, metaheuristic, or reinforcement learning-based. Negotiation outcomes consistently indicate that, regardless of the configuration employed, the use of a KG improves agents’ rewards by at least 1.21 % and up to 90.91 %. Results highlight that the proposed life cycle enables the integration of contextual and domain-specific data and metadata into a KG that enhances agents’ learning, while also allowing for the development and selection of relevant queries and graph data science algorithms to improve the strategic behavior of negotiation agents with cognitive abilities.}, archive = {J_ESWA}, author = {Dan E. Kröhling and Ernesto C. Martínez}, doi = {10.1016/j.eswa.2025.129596}, journal = {Expert Systems with Applications}, month = {3}, pages = {129596}, shortjournal = {Expert Syst. Appl.}, title = {Knowledge graph life cycle for cognitive agents – A case study on automated negotiation in smart grids}, volume = {298}, year = {2026}, }
• (2026). Research on multi-objective optimization of multi-endpoint VRP with time window for the distribution of seasonal products by multi-homing heterogeneous fleets. ESWA, 298, 129595. (www)

  @article{ , abstract = {This paper addresses a novel VRP variant integrating seasonal demand fluctuations, heterogeneous vehicle sources, and multi-endpoint constraints, focusing on the distribution of seasonal products in a steel parts enterprise. It tackles the complex vehicle routing problem with time windows involving heterogeneous fleets, which encompass different vehicle sources (owned and rented), types (fuel-powered and electric), capacities, ranges, and endpoints. To balance enterprise profitability, greenhouse gas emissions, and environmental quality, we develop a mathematical model centered on optimizing distribution costs, greenhouse gas emissions, and vehicle utilization. Drawing inspiration from ancient competitive activities, we propose a novel Huashan Swords Algorithm (HSSA). Through simulations using real enterprise data, we demonstrate the HSSA’s effectiveness, with comparative experiments against existing advanced algorithms highlighting its superiority. Applying the algorithm to design logistics distribution schemes, we conduct in-depth tests considering different customer groups and fuel station distributions. Analyzing the results from the perspectives of profitability, emissions, and environmental quality, we offer targeted operational suggestions for the enterprise based on its situation, geographical characteristics, and fiscal policies. Moreover, we provide recommendations to local governments on fuel station construction and vehicle subsidy policies, contributing practical solutions to both enterprise operations and regional development.}, archive = {J_ESWA}, author = {Zhang Yanhu and Yan Lijuan and Kong ShuMei and Miao Decheng}, doi = {10.1016/j.eswa.2025.129595}, journal = {Expert Systems with Applications}, month = {3}, pages = {129595}, shortjournal = {Expert Syst. Appl.}, title = {Research on multi-objective optimization of multi-endpoint VRP with time window for the distribution of seasonal products by multi-homing heterogeneous fleets}, volume = {298}, year = {2026}, }
• (2026). Revealing the correlation between economic indicators and gold prices for forecasting: Medium term forecast framework with data patch. ESWA, 298, 129594. (www)

  @article{ , abstract = {Predicting gold prices through the analysis of key economic indicators such as inflation rates, Government Bond Yields, and the U.S. Dollar Index, alongside historical Gold Prices, is crucial for enabling investors to better understand market dynamics and make vital decisions to maximize returns. However, previous studies have faced challenges in extracting hidden factors related to gold price prediction from diverse economic indicators, and the comprehensive exploration of gold price data is yet to be fully achieved. To address this, the present study introduces a mid to long-term gold price prediction model named DPformer. This model utilizes a patching strategy to investigate the relationships between different economic indicators and Gold Prices. It also employs a decomposition approach to discover the mid to long-term trend characteristics and yearly seasonal patterns of Gold Prices. The core of the model integrates a Transformer module, which is solely based on an Encoder structure, and enhances it with multiple attention mechanisms and convolutions. This enhancement allows the improved Transformer model to more effectively capture the long-term dependencies of Gold Prices. The empirical results demonstrate that DPformer consistently outperforms a suite of advanced models widely adopted in terms of mid to long-term forecasting accuracy, including LSTM, GRU, Transformer, DLinear, and PatchTST. Notably, for the 30-step gold price prediction task, DPformer achieves a 21.78 % reduction in Mean Squared Error compared to PatchTST. Moreover, by quantitatively analyzing how various economic indicators influence gold price forecasts, this study provides substantial support for investors in making informed decisions at critical moments.}, archive = {J_ESWA}, author = {Guanhao Bao and Yunbo Niu and Baisheng Cui and Wanying Ji}, doi = {10.1016/j.eswa.2025.129594}, journal = {Expert Systems with Applications}, month = {3}, pages = {129594}, shortjournal = {Expert Syst. Appl.}, title = {Revealing the correlation between economic indicators and gold prices for forecasting: Medium term forecast framework with data patch}, volume = {298}, year = {2026}, }
• (2026). AI-driven optimization of project portfolios in corporate ecosystems with synergies and strategic factors. ESWA, 298, 129593. (www)

  @article{ , abstract = {This paper studies the optimization of project portfolios in corporate ecosystems by considering both strategic factors and return synergies between projects. We propose a hybrid method that combines machine learning with mathematical programming to address this enhanced form of project portfolio optimization. Unlike traditional approaches, which evaluate projects mainly based on individual risks and returns, our framework considers strategic priorities and the extra value created when projects reinforce each other. Machine learning models predict synergies, while exact optimization ensures consistent portfolio selection under resource and strategic constraints. A numerical proof-of-concept illustrates the methodology. Computational experiments show that portfolios designed with synergy and strategy in mind might achieve a significantly higher performance than portfolios that do not account for project synergies. The paper also examines computational efficiency and scalability, highlighting the approach’s potential for practical application in complex and dynamic corporate ecosystems.}, archive = {J_ESWA}, author = {Patricia Rodriguez-Garcia and Angel A. Juan and Jon A. Martin and David Lopez-Lopez and Josep M. Marco}, doi = {10.1016/j.eswa.2025.129593}, journal = {Expert Systems with Applications}, month = {3}, pages = {129593}, shortjournal = {Expert Syst. Appl.}, title = {AI-driven optimization of project portfolios in corporate ecosystems with synergies and strategic factors}, volume = {298}, year = {2026}, }
• (2026). Overview of machine learning in class imbalance scenarios: Trends, challenges, and approaches. ESWA, 298, 129592. (www)

  @article{ , abstract = {This study presents a systematic mapping of machine learning in class imbalance scenarios, offering a broad overview of key challenges, promising emerging techniques, and established methodologies across various application domains. The investigation stands out by employing a hybrid search and selection protocol that combines methodological rigor with technical innovation. The adopted strategy integrated manual searches in reference sources with automated processes based on machine learning, semantic embeddings, and graph-based ranking algorithms. To enhance selection quality, the Quasi-Golden Set (QGS) method was used to build a reference set from manually selected articles – a critical foundation for calibrating and evaluating automated search strings. This combination ensured broad coverage of the topic while improving sensitivity and precision in identifying relevant studies. The initial analysis reviewed 25,593 publications. After screening and applying eligibility criteria, 468 articles were included in the final dataset. The results indicate that 55 % of the studies address multiple domains, with a strong predominance of tabular data ( 84 % ). SMOTE and hybrid approaches were among the most common techniques, present in 61 % of the studies. In terms of evaluation metrics, ROC-AUC was the most frequently used, followed by F1-score and accuracy – the latter noted for limitations in highly imbalanced scenarios. Building on these findings, we derive an empirically grounded taxonomy that links problem context, solution algorithms, and scenario-appropriate evaluation metrics, and we provide a minimal selection guideline table to support applied use. While sampling-based methods remain prevalent, deep learning approaches such as convolutional neural networks and graph-based models are increasingly adopted. Additionally, federated, contrastive, and semi-supervised learning are emerging as relevant paradigms, particularly suited for privacy-aware or low-label environments. This study consolidates current knowledge, identifies methodological and application gaps, and highlights trends that are likely to shape future research. It contributes both a comprehensive synthesis of the field and strategic insights for advancing machine learning techniques in the presence of class imbalance.}, archive = {J_ESWA}, author = {Gilberto Sussumu Hida and André Câmara Alves Do Nascimento}, doi = {10.1016/j.eswa.2025.129592}, journal = {Expert Systems with Applications}, month = {3}, pages = {129592}, shortjournal = {Expert Syst. Appl.}, title = {Overview of machine learning in class imbalance scenarios: Trends, challenges, and approaches}, volume = {298}, year = {2026}, }
• (2026). Multi-objective multi-product process planning with reconfigurable machines: Exact and metaheuristic approaches. ESWA, 298, 129591. (www)

  @article{ , abstract = {Process planning in reconfigurable manufacturing systems usually considers a single product, this reduces the efficiency of the overall production plan when multiple products are combined. This paper tackles the Multi-Product Process Planning Problem (MPPP), optimizing both individual process plans and their sequencing. We propose a 0–1 LP model, the model is relaxed by fixing the product sequencing variables and implemented in a Normal-Boundary Intersection method (NBI-es), the method uses a function for iteratively updating β values. Three metaheuristics are also developed: NSGA-II, and two variants of MOEA/D, one enhanced by Opposition-based learning (OBL). Computational experiments show that the update function enhances the performance of NBI over simple Normalized-Weighted Sum (NWS) method. Additionally, NBI-es performs better in HV metric for small size instances if it is given enough CPU times, while MOEA/D significantly outperforms NSGA-II on larger instances on most convergence and spread based metrics. OBL further enhances solution diversity for MOEA/D, albeit with less convergence. A special case of the MPPP is investigated, involving identical products: the Multi-Unit Process Planning (MUPP). An integrated approach was compared with a sequential separated approach. Results indicate that the integrated approach outperforms the separated method for smaller problem instances. Moreover, the analysis of high-quality MUPP solutions revealed a tendency towards diverse process plan combinations rather than repetitive identical ones.}, archive = {J_ESWA}, author = {Abdelkader Mechaacha and Fayçal Belkaid and Nadjib Brahimi}, doi = {10.1016/j.eswa.2025.129591}, journal = {Expert Systems with Applications}, month = {3}, pages = {129591}, shortjournal = {Expert Syst. Appl.}, title = {Multi-objective multi-product process planning with reconfigurable machines: Exact and metaheuristic approaches}, volume = {298}, year = {2026}, }
• (2026). A slack-based two-stage improved particle swarm optimization algorithm for robust scheduling of a flexible job-shop with new and remanufacturing jobs. ESWA, 298, 129590. (www)

  @article{ , abstract = {Remanufacturing has attracted increasing attention for its environmental and economic benefits. Since it is difficult to achieve economies of scale when processing small amounts of remanufacturing jobs alone, these jobs are processed in the same job-shop for new jobs in some enterprises. The processing times of remanufacturing jobs are uncertain due to unpredictable status, leading to certain impacts on scheduling performance. Therefore, we address a flexible job-shop scheduling problem with new and remanufacturing jobs to minimize makespan. To solve this problem, a slack-based two-stage improved particle optimization algorithm is proposed. The first stage aims to yield a solution set with minimum makespan, while the second stage aims to search the best robust solution with maximum total slack from the set. Both stages are executed alternately to optimize makespan and total slack. Moreover, a position updating mechanism with genetic operators and a tabu search inspired local search strategy are implemented to improve algorithmic performance. Computational experiments are conducted using adapted benchmark problems and an industrial case to validate the proposed algorithm.}, archive = {J_ESWA}, author = {Jun Liu and Zhui Gui and An Li and Qiong Liu}, doi = {10.1016/j.eswa.2025.129590}, journal = {Expert Systems with Applications}, month = {3}, pages = {129590}, shortjournal = {Expert Syst. Appl.}, title = {A slack-based two-stage improved particle swarm optimization algorithm for robust scheduling of a flexible job-shop with new and remanufacturing jobs}, volume = {298}, year = {2026}, }
• (2026). Evaluating and predicting green economic efficiency in chinese cities: A three-stage network SBM and machine learning approach. ESWA, 298, 129589. (www)

  @article{ , abstract = {This paper evaluates and predicts green economic efficiency (GEE) across 248 Chinese cities from 2010 to 2021 using a three-stage network SBM model based on subsystems of economic production, social development, and environmental governance. To enhance accuracy in both assessment and forecasting, machine learning methods are incorporated, and the Dagum Gini coefficient is employed to analyze regional disparities. This study innovatively proposes a three-stage network SBM model to resolve the “black box” limitation of conventional DEA approaches, while a DEA-ML model is developed to achieve enhanced prediction accuracy. The results reveal that GEE in Chinese cities remains low, with the eastern region leading and the western region trailing. However, efficiency has improved since 2016, primarily driven by advancements in environmental governance. Regional disparities, largely attributed to interregional differences, are gradually decreasing. Among forecasting models, the backpropagation neural network (BPNN) delivers the highest accuracy, predicting sustained leadership in the east, strong growth in the northeast, and a reduction in national disparities. This study offers a comprehensive framework for evaluating and predicting GEE, providing valuable insights for sustainable development policies.}, archive = {J_ESWA}, author = {Zhishuo Zhang and Hu Liu and Yunpeng Gong and Huayong Niu}, doi = {10.1016/j.eswa.2025.129589}, journal = {Expert Systems with Applications}, month = {3}, pages = {129589}, shortjournal = {Expert Syst. Appl.}, title = {Evaluating and predicting green economic efficiency in chinese cities: A three-stage network SBM and machine learning approach}, volume = {298}, year = {2026}, }
• (2026). HAEA: A heterogeneous alternating evolutionary algorithm for numerical optimization. ESWA, 298, 129587. (www)

  @article{ , abstract = {A hybrid algorithm integrating a couple of individual evolutionary algorithms (sub-algorithms) is widely recognized as an effective approach to enhance both robustness and optimization performance. Nevertheless, such integration often destroys the structure of the sub-algorithm and makes it difficult to incorporate additional evolutionary algorithms. To address these limitations, this study introduces a novel framework, the Heterogeneous Alternating Evolutionary Algorithm (HAEA), designed to integrate multiple evolutionary algorithms while enabling the flexible addition, removal, and replacement of internal sub-algorithms. To facilitate the integration of a broad spectrum of sub-algorithms, this study draws inspiration from the particle swarm optimization algorithm to devise a suite of information indicators for the transmission of optimization information between sub-algorithms with disparate structures. Furthermore, HAEA is endowed with an adaptive mechanism that dynamically modifies the selection probabilities of its sub-algorithms based on their long-term and short-term performance throughout the evolutionary process. We conducted a comparative analysis of HAEA against all its sub-algorithms across three widely recognized function test sets: CEC2013, CEC2017, and CEC2022. Meanwhile, we applied the HAEA separately to basic metaheuristic algorithms and advanced evolutionary algorithms in recent years and conducted two comparative experiments. Both experimental results show that HAEA outperforms all sub-algorithms in terms of robustness and optimization performance. Its distinctive flexibility allows for the incorporation of additional superior evolutionary algorithms in the future, thereby enhancing its overall performance.}, archive = {J_ESWA}, author = {Taiyong Li and Tianhao Yi and Zhenda Hu and Wu Deng and Donglin Zhu and Zhilong Xie and Jiang Wu}, doi = {10.1016/j.eswa.2025.129587}, journal = {Expert Systems with Applications}, month = {3}, pages = {129587}, shortjournal = {Expert Syst. Appl.}, title = {HAEA: A heterogeneous alternating evolutionary algorithm for numerical optimization}, volume = {298}, year = {2026}, }
• (2026). Deep analysis on MLSY for fractional-order higher-dimension-valued neural networks under the action of free quadratic coefficients. ESWA, 298, 129586. (www)

  @article{ , abstract = {In this article, free quadratic coefficients are proposed in order to deeply study the flexible criteria of synchronization problem for two kinds of fractional-order higher-dimension-valued neural networks (FOHDVNN) with usual neurons and threshold ones, respectively. First, the uniform system is constructed for two kinds of FOHDVNN which contains both fractional-order octonion-valued neural networks (FOOVNN) and fractional-order quaternion-valued neural networks (FOQVNN). Based on higher-dimension algebra multiplication rules, the studied FOHDVNN are directly decomposed into the eight or four subsystems in real-valued field. Subsequently, free quadratic coefficients are taken into the establishment of two types of Lyapunov-Krasovskii functional (LKF) which is newer and more general. Then, mainly based on the very recent lemmas and Lyapunov theories, the flexible criteria are generally acquired for the global Mittag-Leffler synchronization (MLSY) problem of FOHDVNN. The final criteria have the advantage in being easily calculated and widely used. It is worth noting that the optimal solutions of these criteria can be obtained through the genetic algorithm and the synchronization performance can be improved by optimizing the quadratic coefficients. Finally, three simulation examples are presented to express the availability and progress of the derived results.}, archive = {J_ESWA}, author = {Jianying Xiao and Yongtao Li}, doi = {10.1016/j.eswa.2025.129586}, journal = {Expert Systems with Applications}, month = {3}, pages = {129586}, shortjournal = {Expert Syst. Appl.}, title = {Deep analysis on MLSY for fractional-order higher-dimension-valued neural networks under the action of free quadratic coefficients}, volume = {298}, year = {2026}, }
• (2026). Complex-order darwinian particle swarm optimization. ESWA, 298, 129584. (www)

  @article{ , abstract = {The Particle Swarm Optimization (PSO) algorithm has been one of the most effective methods for solving various complex optimization problems. However, non-adaptive versions of the PSO do not use historical information for performance enhancement and suffer from performance degradation problems. This paper presents a Complex-Order Darwinian PSO (CoDPSO) algorithm, which effectively enhances the performance of the PSO. A complex-order derivative mechanism is introduced into the velocity update rule to improve local exploitation using historical velocity information. Additionally, a Degradation Elimination (DE) strategy is designed to mitigate performance drop during the optimization process. Sensitivity analysis is conducted to evaluate the impact of control parameters on the algorithm’s behavior, demonstrating its robustness across a wide range of settings. Comparative experiments on CEC 2022 benchmark functions show that the CoDPSO outperforms other PSO variants in terms of accuracy, stability, and convergence. Wilcoxon statistical tests further confirm the significance of these improvements. The experimental results indicate the feasibility and efficiency of the CoDPSO.}, archive = {J_ESWA}, author = {Xiaobo Wu and Liping Chen and Huafeng Li and António M. Lopes and Chuang Liu and Yangquan Chen and Yi Chai}, doi = {10.1016/j.eswa.2025.129584}, journal = {Expert Systems with Applications}, month = {3}, pages = {129584}, shortjournal = {Expert Syst. Appl.}, title = {Complex-order darwinian particle swarm optimization}, volume = {298}, year = {2026}, }
• (2026). A multi-objective framework for predicting public opinion trends on infectious diseases using NSGA-II and interval predictions. ESWA, 298, 129583. (www)

  @article{ , abstract = {Predicting public opinion trends during major infectious disease outbreaks is critical for guiding effective public health responses. However, predicting public opinion remains challenging because it is influenced by socio-economic, psychological, and media factors. This paper presents a novel framework for predicting public opinion trends related to significant infectious diseases, with a focus on COVID-19 as a case study. The proposed framework identifies the key factors influencing public opinion development and enables both point and interval predictions. The framework uses information ecology theory and applies the NSGA-II algorithm to select the features that best drive public opinion trends. By incorporating this framework, accurate point forecasts are produced alongside prediction intervals, effectively quantifying the uncertainty inherent in public opinion dynamics. This approach minimizes the quality-driven loss function to generate precise prediction intervals, providing decision-makers with critical insights into public opinion fluctuations during epidemics. The results offer valuable, real-time public sentiment warnings, supporting timely and effective interventions in epidemic prevention and control efforts.}, archive = {J_ESWA}, author = {Futian Weng and Meng Su and Petr Hajek and Mohammad Zoynul Abedin}, doi = {10.1016/j.eswa.2025.129583}, journal = {Expert Systems with Applications}, month = {3}, pages = {129583}, shortjournal = {Expert Syst. Appl.}, title = {A multi-objective framework for predicting public opinion trends on infectious diseases using NSGA-II and interval predictions}, volume = {298}, year = {2026}, }
• (2026). A deep reinforcement learning-guided multimodal multi-objective evolutionary algorithm with a serial-parallel mechanism. ESWA, 298, 129581. (www)

  @article{ , abstract = {The core challenge for multimodal multi-objective problem (MMOP) resolution lies in maintaining synergistic interactions between convergence and diversity. However, the existing algorithms usually consider convergence-first, which neglect to consider both diversity and convergence into account during the evolutionary process. Likewise, the optimization methods tend to gravitate toward locally optimal regions rapidly, leading to lose diversity for the local PS. This paper proposes a Deep Reinforcement Learning-guided multimodal multi-objective evolutionary algorithm with a serial-parallel mechanism (DRLMMEA) to investigate the impact of different operator selection on the performance of MMEAs, which greatly helps to balance the convergence and diversity. DRLMMEA utilizes Q-Network to select the operator with the highest reward to enhance the population’s search ability. An improved sorting method (ISM) based on neighborhood dominance updates the population by sorting individuals according to their convergence quality, thereby enhancing convergence performance in the objective space. Moreover, this study proposes a series-parallel mechanism, a series structure enhances the diversity in the decision space, while the parallel structure reduces the computational burden of the algorithm evidently. The proposed Deep Reinforcement Learning-assisted operator selection mechanism, which enables effective balance between diversity and convergence, and an improved crowding distance approach that enhances convergence performance. DRLMMEA undergoes comprehensive testing against 6 contemporary approaches using MMF and IDMP benchmark problems, achieving supremacy in 4 principal performance metrics according to experimental findings. The multimodal gearbox parameter optimization is addressed using the proposed DRLMMEA, which demonstrates superior performance against 6 algorithms in comparative evaluations. It has demonstrated a significant role in solving the MMOPs with the imbalance between convergence and diversity.}, archive = {J_ESWA}, author = {Ying Huang and Xiaojian Cao and Benben Zhou and Wei Li and Shuling Yang and S.M. Shafi and Zhou Yang}, doi = {10.1016/j.eswa.2025.129581}, journal = {Expert Systems with Applications}, month = {3}, pages = {129581}, shortjournal = {Expert Syst. Appl.}, title = {A deep reinforcement learning-guided multimodal multi-objective evolutionary algorithm with a serial-parallel mechanism}, volume = {298}, year = {2026}, }
• (2026). Evolutionary multitasking optimization based on cross-task association mapping strategy. ESWA, 298, 129580. (www)

  @article{ , abstract = {In multitasking optimization, knowledge transfer between tasks through subspace generation has been widely employed to enhance the convergence performance of algorithms. However, this approach fails to account for the inter-task knowledge mapping relationships. Therefore, cross-task knowledge transfer during the optimization process remains inherently blind, potentially leading to mismatched subspace information and consequently degrading the algorithm’s performance. To address this issue, this paper proposes a multitask evolutionary algorithm based on an association mapping strategy and an adaptive population reuse mechanism, namely PA-MTEA. Specifically, to fully represent the correlations between multitask domains and enhance the adaptability of transfer solutions in target tasks, this paper introduces a subspace projection strategy based on partial least squares, which achieves the correlation mapping between the source and target tasks during the dimensionality reduction of the search space. Additionally, to further enhance knowledge transfer across tasks, an alignment matrix is obtained by adjusting the subspace Bregman divergence after deriving the respective subspaces, minimizing variability between task domains. Finally, to balance the global exploration of algorithms with local exploitation, an adaptive population reuse mechanism based on the residual structure is designed. This mechanism reuses historically successful individuals to guide the evolutionary direction of the population, thus improving the algorithm’s convergence performance. Experimental results on various benchmark suites and real-world cases demonstrate that PA-MTEA exhibits significantly superior performance compared to six other advanced multitask optimization algorithms.}, archive = {J_ESWA}, author = {Tao Yin and Lizhong Yao and Xin Zong and Pengjie Qin and Haoming Dong}, doi = {10.1016/j.eswa.2025.129580}, journal = {Expert Systems with Applications}, month = {3}, pages = {129580}, shortjournal = {Expert Syst. Appl.}, title = {Evolutionary multitasking optimization based on cross-task association mapping strategy}, volume = {298}, year = {2026}, }
• (2026). UniTask+: Exploring and unifying strong and weak task-aware consistency for semi-supervised blastocyst image segmentation. ESWA, 298, 129578. (www)

  @article{ , abstract = {The segmentation of different tissues within blastocysts is essential for embryologists to objectively observe and evaluate embryos, thereby contributing to a higher success rate of in vitro fertilization treatment. Inspired by the primary observation of skeletal patterns and boundary information by clinical doctors, we present an interesting task-aware view for blastocyst segmentation with semi-supervised learning, focusing on task-invariant and task-specific dependencies of segmentation. Firstly, we explore one strong-correlation task with bidirectional transformation between its outputs and the segmentation results, and another weak-correlation task with monodirectional transformation from segmentation maps. The correlation among different tasks inspires us to propose Task-Aware Smoothness (TAS) Assumption , thereby deducing different types of task-aware consistency. Then, a new Unified Task-aware Consistency Interaction (UniTask+) framework is developed to unify and fully take advantage of these strong, weak, and strong-to-weak task-aware consistency. It is comprised of a medical segmentation (MS) branch to implement segmentation and two extra branches performing strong/weak-correlation tasks based on the same backbone. Concretely, a level-set (LS) branch promotes the strong consistency while a point-set (PS) branch stimulates the weak consistency with underlying task perturbations. Numerous experiments have been conducted on the inner cell mass (ICM), blastocyst, proving the effectiveness of our tactics. Furthermore, we have also conducted experiments with datasets from the left atrium (LA), which shares similar structural features with embryos, to validate the robustness of the model. Our methods have shown prominent improvements over up-to-date SSL methods, which advocates our precedent hypothesis.}, archive = {J_ESWA}, author = {Hua Wang and Linwei Qiu and Jingfei Hu and Jicong Zhang}, doi = {10.1016/j.eswa.2025.129578}, journal = {Expert Systems with Applications}, month = {3}, pages = {129578}, shortjournal = {Expert Syst. Appl.}, title = {UniTask+: Exploring and unifying strong and weak task-aware consistency for semi-supervised blastocyst image segmentation}, volume = {298}, year = {2026}, }
• (2026). Deep asymmetric semantic hashing with probability shifting for multi-label image retrieval. ESWA, 298, 129577. (www)

  @article{ , abstract = {Deep hashing has been widely used in large-scale multimedia retrieval due to its advantages in terms of low storage cost and computational efficiency. Deep hashing algorithms can jointly learn semantic features and hash functions, encoding the original data into compact binary codes with significant discriminative power. However, in multi-label scenarios, especially when the number of samples is extremely large, a high negative-positive imbalance may occur, particularly when the proportion of negative samples is too high, which can lead to bias in the semantic relationships between the learned images. To solve this problem, symmetric losses such as focal loss were proposed, which treat positive and negative samples equally, but the retrieval results are suboptimal. This may be because the equal-weighted processing strategy causes the model to over-focus on hard negative samples and ignore the learning of positive sample features. Besides, mislabeled negative samples, especially those with a probability close to 1, can lead the model to learn incorrect features, harming its discrimination ability, reducing accuracy and recall, and causing overfitting and poor generalization. Accordingly, this paper proposes a novel hashing model, Deep Asymmetric Semantic Hashing with Probability Shifting framework (DASH-PS), for discriminative binary code learning. Specifically, by combining asymmetric focusing strategy and probability shifting strategy, asymmetric semantic loss is designed to solve negative-positive imbalance and ground-truth mislabeling. To keep the contribution of positive samples while focusing on hard negative samples, asymmetric focusing strategy is proposed to decouple negative and positive samples and assign different attenuation factors. By offsetting the probability of negative samples, probability shifting strategy completely discards easy negative samples and very hard negative samples suspected of being mislabeled. Additionally, an adaptive asymmetric learning mechanism is proposed to reduce the fixed difference in average probabilities between positive and negative samples, thereby simplifying hyperparameter selection and improving retrieval efficiency. Extensive experimental results on multiple benchmark datasets validate that our DASH-PS outperforms various state-of-the-art hashing methods. The code for the implementation of our DASH-PS framework is available at https://github.com/QinLab-WFU/DASH-PS.}, archive = {J_ESWA}, author = {Yongyue Fu and Qibing Qin and Jinkui Hou and Congcong Zhu and Lei Huang and Wenfeng Zhang}, doi = {10.1016/j.eswa.2025.129577}, journal = {Expert Systems with Applications}, month = {3}, pages = {129577}, shortjournal = {Expert Syst. Appl.}, title = {Deep asymmetric semantic hashing with probability shifting for multi-label image retrieval}, volume = {298}, year = {2026}, }
• (2026). Physics-informed tensor autoencoder with memory for video anomaly detection. ESWA, 298, 129576. (www)

  @article{ , abstract = {Video data can be naturally represented as tensors. Despite great progress in anomaly detection with memory-augmented autoencoders, the memory module therein can only handle vectors and inevitably breaks tensor structures, thus leading to performance degradation. Moreover, after the mapping of the encoder, some abnormal features may directly fall into the normal convex polytope, as autoencoders only use the output error to guide the construction of latent variables without imposing any constraint. The memory module can not handle these abnormal features, so that the abnormal observations may not be identified. To solve these problems, we propose a Physics-Informed Tensor AutoEncoder (PITAE) framework, which incorporates both neural networks and physical laws, i.e., tensor operation rules. Specifically, we design a tensor decomposition network followed by an explicit tensor operation to decompose the latent variable into low-rank and sparse components, and only the low-rank component is inputted to the decoder. In this way, we reserve the tensor structure and meanwhile impose a low-rank constraint on the latent variable, thereby compressing the features of normal samples into a ”smaller” region where anomalies are less likely to fall into. Consequently, the non-low-rank anomalies can be identified. But the low-rank anomalies may still not be identified. To further solve this problem, we design a tensor Memory module, and the overall model is named as PITAEM. Finally, based on the proposed framework, we design a novel composite anomaly score to identify anomalies of various kinds. Experiments on various video datasets demonstrate the effectiveness of the proposed method, especially in the small data regime.}, archive = {J_ESWA}, author = {Jianan Liu and Chunguang Li}, doi = {10.1016/j.eswa.2025.129576}, journal = {Expert Systems with Applications}, month = {3}, pages = {129576}, shortjournal = {Expert Syst. Appl.}, title = {Physics-informed tensor autoencoder with memory for video anomaly detection}, volume = {298}, year = {2026}, }
• (2026). A robust medical image encryption technique using inverse cosine chaotic map. ESWA, 298, 129574. (www)

  @article{ , abstract = {With the rapid proliferation of digital imaging technologies, the need for robust and lightweight image encryption techniques has become increasingly critical, particularly for medical, military, and personal data applications. In this paper, we propose a novel image encryption scheme based on a one-dimensional inverse cosine chaotic map (1D-ICC), which introduces a highly sensitive and structurally complex nonlinear dynamical system. The proposed method integrates a dynamic Josephus-based intra-block scrambling mechanism, a global zigzag permutation strategy, and an adaptive diffusion process guided by chaotic sequences, thereby enhancing the confusion and diffusion characteristics of the cipher. Unlike conventional approaches, our scheme dynamically derives the encryption key from the SHA-512 hash of the original image, ensuring both sensitivity to plaintext changes and resistance to known-plaintext and chosen-plaintext attacks. The use of the 1D-ICC map, featuring a tunable control parameter r 5 enables rich chaotic behavior even in one dimension, reducing computational complexity without sacrificing security. Comprehensive experiments validate the robustness and efficiency of the encryption scheme, with performance metrics including correlation coefficients below 0.003, information entropy of 7.9993, a Number of Pixels Change Rate (NPCR) of 99.61 %, and a Unified Averaged Changed Intensity (UACI) of 33.42 %. These results demonstrate that our method surpasses several existing techniques in both security strength and computational performance, underscoring the potential of the 1D-ICC map for practical image encryption applications.}, archive = {J_ESWA}, author = {Jackson J and Perumal R}, doi = {10.1016/j.eswa.2025.129574}, journal = {Expert Systems with Applications}, month = {3}, pages = {129574}, shortjournal = {Expert Syst. Appl.}, title = {A robust medical image encryption technique using inverse cosine chaotic map}, volume = {298}, year = {2026}, }
• (2026). Enhancing text classification with neural label embedding and weakly-supervised learning. ESWA, 298, 129569. (www)

  @article{ , abstract = {Recent years have witnessed the widespread adoption of deep-learning-based models in a range of linguistic tasks including the fundamental text classification. These deep neural networks, however, often face challenges due to the limited availability of large-scale training data with high-quality label annotations. Furthermore, while supervised learning has proven to be superior in training sentence representations for downstream tasks like text classification, this aspect has received relatively little attention. In this study, a novel model named L abel Em bedding joint with We akly-supervised C lassification ( LemWec ) is proposed, which aims to establish a unified framework by combining supervised sentence embedding with multiclass classification. For supervised sentence embeddings, the model incorporates seed information such as label names and designs an encoder network with a new pooling layer. Additionally, the model adopts a pseudo-labeling approach to leverage a substantial amount of unlabeled samples. This approach specifically addresses the drawback of generating pseudo-labels with the highest confidence and introduces a noise adaptation method to mitigate this issue. The results of extensive experiments conducted on four real-world datasets demonstrate that the proposed LemWec model can significantly enhance the performance of text classification when compared to a comprehensive set of baselines.}, archive = {J_ESWA}, author = {Xiao Jing and Zhe Li and Zhiang Wu and Dejun Mu}, doi = {10.1016/j.eswa.2025.129569}, journal = {Expert Systems with Applications}, month = {3}, pages = {129569}, shortjournal = {Expert Syst. Appl.}, title = {Enhancing text classification with neural label embedding and weakly-supervised learning}, volume = {298}, year = {2026}, }
• (2026). Self-supervised ZINB-based kernel representation learning of single-cell RNA sequencing data. ESWA, 298, 129568. (www)

  @article{ , abstract = {Cell clustering plays a vital role in single-cell RNA sequencing (scRNA-seq) data analysis. Although many deep cell clustering methods have been proposed to cluster the scRNA-seq data, they overlook the structural partitioning objectives during the representation learning process, leading to challenges with non-linearly separable structures. In this paper, we present a novel end-to-end deep kernel cell clustering model for scRNA-seq data based on self-supervised ZINB-based kernel representation learning, named scDKC, which simultaneously learns cell kernel representations and identifies cell clusters. Specifically, a kernel-aid hybrid representation learning encoder is developed to effectively learn the separable kernel representation of cells, consisting of cells’ expression characteristics and cell-cell topological interactions. To guide the direction of kernel representation learning, a ZINB-based kernel representation learning decoder is designed by capturing the global probabilistic structure, the representation and the cell graph structure of the scRNA-seq data. By leveraging the clustering self-supervised strategy, representation self-supervised strategy, ZINB-based distribution self-supervised strategy, and kernel self-supervised strategy, scDKC optimizes cell cluster label assignment and learns cell kernel representations through a joint mutual self-supervised mechanism. Extensive experiments on 15 real scRNA-seq datasets, comparing scDKC with 10 competing methods, highlight its competitive advantages.}, archive = {J_ESWA}, author = {Lina Ren and Maoxuan Yao and Ruizhang Huang and Yongbin Qin}, doi = {10.1016/j.eswa.2025.129568}, journal = {Expert Systems with Applications}, month = {3}, pages = {129568}, shortjournal = {Expert Syst. Appl.}, title = {Self-supervised ZINB-based kernel representation learning of single-cell RNA sequencing data}, volume = {298}, year = {2026}, }
• (2026). A dual uncertainty-aware fusion framework for face expression recognition in the wild. ESWA, 298, 129567. (www)

  @article{ , abstract = {Facial Expression Recognition(FER) is a key task in the broader landscape of affective computing and human-computer interaction, enabling machines to interpret human emotions. To better learn discriminative features under complex facial variations, recent FER research has increasingly adopted multi-branch fusion architectures that aim to capture complementary features from diverse perspectives. However, existing multi-branch fusion strategies, including static weighting, simple concatenation, or uncertainty-aware modeling, lack the capacity to comprehensively capture and reconcile the reliability variations across both individual instances and structural branches. To overcome these limitations, we propose a novel multi-branch fusion strategy, named Dual Uncertainty-Aware Fusion Framework(DUAFF), which improves the discriminability of integrated features by simultaneously modeling instance-wise uncertainty and inter-branch correlations. Specifically, the proposed method comprises two complementary modules: Instance-Discrepant Uncertainty-Aware Fusion Module (ID-UAFM) and Branch-Discrepant Uncertainty-Aware Fusion Module (BD-UAFM). ID-UAFM is introduced to perform channel-wise entropy analysis between semantically distinct samples to estimate instance-level uncertainty, enabling selective channel-wise fusion that emphasizes reliable representations while suppressing uncertain responses. BD-UAFM is further proposed to capture structural uncertainty by evaluating the relative reliability of features across multiple branches and adaptively weighting their contributions based on inter-branch discrepancies. Experimental results demonstrate that the proposed DUAFF consistently outperforms POSTER across three benchmark datasets, achieving accuracy improvements of 0.23 % on RAF-DB, 0.69 % on FER2013, and 0.29 % on AffectNet (7-class), thereby confirming its effectiveness in enhancing the reliability and discriminability of facial representations.}, archive = {J_ESWA}, author = {Wenfeng Jiang and Ziyi Zhao and Lin Wang and Fang Liu and Chunmei Qing and Xiaofen Xing and Xiangmin Xu and Weiquan Fan and Zhanpeng Jin}, doi = {10.1016/j.eswa.2025.129567}, journal = {Expert Systems with Applications}, month = {3}, pages = {129567}, shortjournal = {Expert Syst. Appl.}, title = {A dual uncertainty-aware fusion framework for face expression recognition in the wild}, volume = {298}, year = {2026}, }
• (2026). Forecasting tourism stock index dynamics: A multiscale deep learning framework integrating emerging media data. ESWA, 298, 129566. (www)

  @article{ , abstract = {This study introduces a novel multiscale and multivariable deep learning framework for tourism stock index forecasting. To address the research gap concerning emerging media’s impact on the tourism sector, our study innovatively integrate multi-source data, including Douyin (China’s prominent short video platform), into our predictive model. Our methodology employs a multiscale decomposition strategy to streamline feature extraction complexity, coupled with an enhanced temporal convolutional network model incorporating soft-thresholding denoising to mitigate noise interference. Furthermore, we implement an adaptive differentiated prediction strategy to optimize model flexibility. Empirical analysis utilizing the CSI Tourism Stock Index demonstrates that our proposed model outperforms benchmark models in both predictive accuracy and stability, thereby validating its efficacy in tourism stock index forecasting.}, archive = {J_ESWA}, author = {Feng Shen and Shuai Huang and Wanqing Zhao and Dao Lan}, doi = {10.1016/j.eswa.2025.129566}, journal = {Expert Systems with Applications}, month = {3}, pages = {129566}, shortjournal = {Expert Syst. Appl.}, title = {Forecasting tourism stock index dynamics: A multiscale deep learning framework integrating emerging media data}, volume = {298}, year = {2026}, }
• (2026). Enhancing adversarial transferability through frequency-domain boundary samples tuning. ESWA, 298, 129565. (www)

  @article{ , abstract = {Transfer-based attacks evaluate the robustness of deep learning models and advance adversarial research to improve the security and reliability of deep learning and its applications. Previous efforts have improved the transferability through advanced gradients, augmented models, or augmented data. In this paper, we understand and enhance the transferability from a new perspective. Delving into intermediate features, we empirically find a difference between the distances of adversarial and original samples from cluster centers of the original classes. The adversarial samples are simultaneously far from both the original samples and the cluster centers, close to generalized decision boundaries. Based on this observation, we propose a novel spectrum tuning attack. Boundary samples are utilized to guide the generation of adversarial samples that are far away from the cluster centers. Specifically, randomized boundary samples are generated by frequency-domain transformations. With the gradients of the diverse boundary samples, the adversarial perturbation moves the examples away from cluster centers, thus approaching generalized decision boundaries. In the optimization process, conjugate directions are employed to avoid oscillations and stabilize the update direction. Given the strong Wolfe parameters, the analysis of the descent direction and current gradient further ensures the convergence speed and stability of the optimization. In addition, Gaussian preprocessing is introduced to smooth the update direction, further stabilize the direction and enhance the transferability. The proposed method is flexible enough to be combined with existing methods to further improve the transferability. Experiments conducted on the ImageNet-compatible dataset validate the effectiveness of the proposed method, e.g., 92.6 % success rate against nine defense methods.}, archive = {J_ESWA}, author = {Shuyan Cheng and Peng Li and Keji Han and Yangjun Xiong and He Xu and Ruchuan Wang}, doi = {10.1016/j.eswa.2025.129565}, journal = {Expert Systems with Applications}, month = {3}, pages = {129565}, shortjournal = {Expert Syst. Appl.}, title = {Enhancing adversarial transferability through frequency-domain boundary samples tuning}, volume = {298}, year = {2026}, }
• (2026). When graph anomaly breaks the coherence: A multi-evidence approach with language models. ESWA, 298, 129557. (www)

  @article{ , abstract = {Graph anomaly detection is critical in domains such as healthcare and economics, where identifying deviations can prevent substantial losses. However, current detection methods, primarily reliant on Graph Neural Networks (GNNs), suffer from a critical limitation: they make judgment on a single piece of evidence – the classification of learned node representations. This “single-verdict” approach is inherently susceptible to misjudgments arising from noisy or biased representations. To address this limitation, we introduce Multi-AD, a novel Multi-evidence-based graph Anomaly Detection framework that leverages the power of Language Models (LMs) to enable more robust and reliable anomaly detection. We provide a paradigm shift by constructing multiple evidence sequences for each target node, and employing LMs to assess the coherence of these sequences. By aggregating coherence scores across multiple sequences, Multi-AD leverages converging evidence to make more informed decisions about anomaly status as the presence of anomalous nodes disrupts coherence. Furthermore, we introduce a coherence-aware edge representation method to enhance the discriminative power of the constructed sequences and a multi-round adaptive integration strategy to handle challenging scenarios where normal nodes might be surrounded by anomalies. Extensive experiments demonstrate that Multi-AD consistently outperforms state-of-the-art methods.}, archive = {J_ESWA}, author = {Xuan Cheng and Jiahui Lu and Chunjing Xiao and Meiyi Yang and Meihui Zhong and Fan Zhou}, doi = {10.1016/j.eswa.2025.129557}, journal = {Expert Systems with Applications}, month = {3}, pages = {129557}, shortjournal = {Expert Syst. Appl.}, title = {When graph anomaly breaks the coherence: A multi-evidence approach with language models}, volume = {298}, year = {2026}, }
• (2026). Multimodal joint subspace model for parkinson’s disease diagnosis. ESWA, 298, 129556. (www)

  @article{ , abstract = {Parkinson’s disease (PD) is an irreversible neurodegenerative disorder that significantly impacts patients’ lives. Accurate early diagnosis prediction is crucial for providing timely treatment to delay disease progression. However, current diagnostic methods predominantly rely on the experience and judgment of clinicians, introducing subjectivity and a lack of standardized, quantitative measures. Sparse subspace learning, as a machine learning technique, can extract critical information from multimodal data while addressing issues such as noise, high-dimensional complexity, and class imbalance. Our study utilizes longitudinal, multimodal neuroimaging data collected at multiple time points to develop a diagnostic model for PD. The approach involves extracting latent local features and leveraging deep learning techniques to generate a comprehensive global feature subset. Adaptive sparse selection is employed to reduce feature redundancy. Finally, support vector machine is used for classification and regression tasks, specifically for PD diagnosis and disease progression score prediction. Extensive experiments were conducted on the PPMI dataset, achieving an accuracy of 90.78 % for Scan Without Evidence of Dopaminergic Deficit (SWEDD) vs. Normal Control (NC) classification, 83.79 % for PD vs. NC, and 91.50 % for PD vs. SWEDD. The results demonstrate that the proposed method improves PD classification and prediction performance, showing promise for early diagnostic applications.}, archive = {J_ESWA}, author = {Haojie Song and Haijun Lei and Yukang Lei and Zhongwei Huang and Jiaqiang Li and Tianfu Wang and Peng Yang and Baiying Lei}, doi = {10.1016/j.eswa.2025.129556}, journal = {Expert Systems with Applications}, month = {3}, pages = {129556}, shortjournal = {Expert Syst. Appl.}, title = {Multimodal joint subspace model for parkinson’s disease diagnosis}, volume = {298}, year = {2026}, }
• (2026). A spectral relevance analysis approach to pattern recognition of financial time series. ESWA, 298, 129555. (www)

  @article{ , abstract = {Understanding patterns in financial time series is crucial for improving prediction accuracy in algorithmic trading and risk management. This paper presents a novel AI-based computer vision approach for classifying financial time series. Historical price sequences are transformed into Gramian Angular Difference Field (GADF) images and fed into a convolutional neural network (CNN) for pattern recognition. To interpret the CNN’s decision-making process, we apply Spectral Relevance Analysis (SpRAy), enabling the identification of distinct clusters based on relevance maps. Clustering the images according to their relevance profiles reveals groups with significantly higher predictive performance compared to the full dataset. The corresponding relevance patterns highlight favorable price movement structures and are identified via the associated clusters.}, archive = {J_ESWA}, author = {Christine Distler and Yarema Okhrin and Jonathan Pfahler}, doi = {10.1016/j.eswa.2025.129555}, journal = {Expert Systems with Applications}, month = {3}, pages = {129555}, shortjournal = {Expert Syst. Appl.}, title = {A spectral relevance analysis approach to pattern recognition of financial time series}, volume = {298}, year = {2026}, }
• (2026). A novel object detection system for computer night vision images using residual 3D transformer-based YoloV8 with adaptive GRU in edge and cloud sector. ESWA, 298, 129554. (www)

  @article{ , abstract = {Object identification is one of the computer vision-based methods used in locating and labelling objects in images. Object detection has been greatly advanced as it is now applicable for detecting night vision images with great accuracy. Most accurate object detection at night can be useful in many applications like nighttime driving, regulating harsh traffic in harsh weather conditions, and surveillance. Object detection in normal conditions can be smoother, but low illumination and harsh weather can lead to low versatility. Images captured at night can reflect a lot of noise with low visual features. Due to its challenging nature, a highly effective object detection model is a challenge for high-level applications. Traditional models still face issues and challenges related to uneven light conditions, brightness variations, different light sources, and noisy backgrounds that need to be addressed. Thus, it is necessary to develop an object detection model for dealing with images with low illumination and varying light conditions. Hence, in this work, an effective object detection framework is implemented for night vision images. At first, from the standard datasets, the significant night vision images are fetched and fed into the proposed model as a Residual 3D Transformer-based YoloV8 with an Adaptive Gated Recurrent Unit (R3DT-YAGRU) for detecting the objects. This includes combining spatial–temporal modelling capabilities into YOLOv8, particularly using 3D transformers for improved feature extraction and an adaptive GRU to manage temporal dependencies at night. Here, the Modified Random Variable-based Dollmaker Optimization Algorithm (MRV-DOA), which is a metaheuristic algorithm motivated by the doll-making procedure. Also, it helps in balancing the exploration phase and exploitation phase to discover the best solutions and is used for tuning the parameters of the R3DT-YAGRU model. At last, the experimental validation is carried out for the recommended object detection process by comparing with other models to establish the supremacy of the suggested work. From the study, the suggested framework achieves an accuracy of 96%, leading to enhanced decision making and better accuracy than other conventional models. The work has prepared its implementation accessible at https://github.com/charlesvprabhu56/Object-Detection .}, archive = {J_ESWA}, author = {V.Charles Prabu and Queen Mary Vidya. M and V. Sathiyamoorthi and P. Durgadevi and M. Gowthami}, doi = {10.1016/j.eswa.2025.129554}, journal = {Expert Systems with Applications}, month = {3}, pages = {129554}, shortjournal = {Expert Syst. Appl.}, title = {A novel object detection system for computer night vision images using residual 3D transformer-based YoloV8 with adaptive GRU in edge and cloud sector}, volume = {298}, year = {2026}, }
• (2026). Learning diversified features for pulmonary hypertension detection using chest X-ray. ESWA, 298, 129553. (www)

  @article{ , abstract = {Compared to traditional Computed Tomography (CT) scans and floatation catheters, chest X-ray offers an efficient, safe and timely examination paradigm, with broader range of scenarios (including intensive care units), for the detection of Pulmonary Arterial Hypertension (PAH). However, it is difficult to learn the variable radiological features of PAH from X-rays due to its low resolution and low contrast. To address the above issues, we propose a diversified features learning framework to fully explore the PAH-related representation from chest X-ray. We first employ a Chest Feature Enhancement Attention (CFEA) module to enhance the initial feature representation. Then, we employ the Deep Temporal Anti-Interference Metric Learning (TAIML) module to fully explore the PAH-related features. We incorporate the information on the temporal evolution of patients’ conditions. Specifically, a patient x , after undergoing treatment, may exhibit two possible states: x + (ill) and x − (cured). Therefore, we can define the distance d ( x , x + ) as the intra-class structural distance, and the distance d ( x , x − ) as the inter-class safe distance. Unlike existing metric learning, we adopt a new strategy: we push positive samples towards negative samples, but ensure distance between them is no less than d ( x , x − ) , thereby enhancing intra-class diversity while maintaining discriminability. Meanwhile, we ensure that the distance between positive samples is greater than d ( x , x + ) , thereby preserving the intra-class structure. Through these two steps, we can learn a diversified but discriminative representation of PAH. Comprehensive experiments showed the our model achieved an impressive accuracy of 86.27 % and an AUC of 0.857 in identifying PAH patients. The code is available at https://github.com/zgfdmn/PAH .}, archive = {J_ESWA}, author = {Chengjin Yu and Huanghui Wang and Yuanting Yan and Zhuyang Chu and Dongsheng Ruan}, doi = {10.1016/j.eswa.2025.129553}, journal = {Expert Systems with Applications}, month = {3}, pages = {129553}, shortjournal = {Expert Syst. Appl.}, title = {Learning diversified features for pulmonary hypertension detection using chest X-ray}, volume = {298}, year = {2026}, }
• (2026). TriHAtt-BERT-DBiLSTM: Deep learning model for medical emergency prediction from the unstructured data. ESWA, 298, 129552. (www)

  @article{ , abstract = {In recent years, the emergency triage has faced several challenges, including insufficient manual triage with physicians, limited medical resources contributing to incorrect triage, overcrowding in the emergency department (ED), and extended patient waiting time. Hence, the Medical Emergency prediction remains the major research area that identifies emergencies about specific diseases using the Medical Transcriptions (MT) provided by physicians. However, the existing methods face the challenges of handling the ambiguity of words, unstructured data, and increased computation complexity. Consequently, this research proposes the Tri-Head Attention-based Bidirectional Encoder Representations from Transformers enabled Distributed Bidirectional Long-Short Term Memory (TriHAtt-BERT-DBiLSTM) for predicting medical emergencies. Specifically, the proposed approach integrates the Tri-Head Attention mechanisms into BERT, which is further hybridized with the DBiLSTM model that offers the synergic strength of providing the dense feature representations to capture the complex dependencies, and enhancement of model ability with the structured parameters to facilitate the medical emergency prediction. Besides, the utilization of BERT in the proposed approach assists in capturing more complex language representations and further executes a better embedding representation of words. The TriHAtt-BERT-DBiLSTM model surpasses other state-of-the-art techniques and achieves 96.40% of accuracy, 96.12% of F1-score, 96.09% of precision, and 96.15% of recall for medical emergency prediction.}, archive = {J_ESWA}, author = {Amita Mishra and Sunita Soni}, doi = {10.1016/j.eswa.2025.129552}, journal = {Expert Systems with Applications}, month = {3}, pages = {129552}, shortjournal = {Expert Syst. Appl.}, title = {TriHAtt-BERT-DBiLSTM: Deep learning model for medical emergency prediction from the unstructured data}, volume = {298}, year = {2026}, }
• (2026). System-level integration of deep learning and computer vision for contact ring seal defect detection in semiconductor manufacturing. ESWA, 298, 129551. (www)

  @article{ , abstract = {The contact ring seals (CRSs) used in electroplating processes during semiconductor manufacturing are susceptible to degradation through chemical etching, electrochemical dissolution, and mechanical wear mechanisms. Despite the implementation of state-of-the-art surface treatment and coating technologies to mitigate CRS corrosion, manual intervention remains frequently required to address this problem. Conventional static defect detection systems for CRSs rely on predefined regions of interest (ROIs) and threshold-based defect area calculations, with surface anomalies identified by comparing the percentage of defective areas within these ROIs. However, this approach exhibits detection failures for millimeter-scale defects, low-contrast anomalies, and geometrically irregular patterns, especially under complex or dynamic environmental conditions. To address these systematic detection failures, we developed a dynamic defect detection system for CRSs by integrating artificial intelligence and traditional computer vision algorithms, achieving a 5.2x improvement in defect detection sensitivity. This system achieved detection accuracy and recall values of over 99 % as well as a response time of 1.43 s average latency, thereby demonstrating a substantial performance improvement compared to a static system, which achieved a recall rate of 18.9 % on the adopted dataset. The system satisfies real-time processing requirements while substantially reducing the need for manual intervention in defect detection and increases production efficiency. Finally, the experimental results of this study indicated that the postprocessing approaches used in the developed system enabled it to flexibly adapt to the different requirements of various production environments.}, archive = {J_ESWA}, author = {Ting-Han Chen and Hsin-Hung Chou and Shuang Zou and Yu-Han Chen and Sun-Yuan Hsieh}, doi = {10.1016/j.eswa.2025.129551}, journal = {Expert Systems with Applications}, month = {3}, pages = {129551}, shortjournal = {Expert Syst. Appl.}, title = {System-level integration of deep learning and computer vision for contact ring seal defect detection in semiconductor manufacturing}, volume = {298}, year = {2026}, }
• (2026). Complementary information-guided interactive fusion network for HSI and LiDAR data joint classification. ESWA, 298, 129549. (www)

  @article{ , abstract = {The classification of land remote sensing using single-modal data has reached a bottleneck, which has spurred significant interest in the joint utilization of multimodal remote sensing data to enhance classification performance. However, existing methods exhibit limitations in extracting intricate local and global features. Furthermore, achieving effective information interaction and deep fusion between multimodal datasets remains an unresolved challenge. To address these issues, we propose a Complementary Information-Guided Interactive Fusion Network (CIGIF-Net) for the classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) data. The core idea of our approach leverages the capability of Convolutional Neural Networks (CNNs) to extract local spatial features while utilizing the strengths of Transformers in modeling long-range dependencies. Furthermore, our method facilitates deep fusion by designing mechanisms for the interactive integration of multimodal local spatial features, complemented by guidance from multimodal data during long-range dependency modeling, thereby improving overall classification performance. Specifically, CIGIF-Net incorporates multiscale feature learning, interactive feature fusion, and the complementary information-guided attention mechanism. Initially, CNNs are used to learn multiscale local spatial features. Subsequently, we perform an interactive fusion of multimodal spatial information based on channel attention techniques. Finally, the complementary information-guided attention mechanism dynamically utilizes complementary insights to inform deeper attention distributions, which guide global feature construction and enable efficient information aggregation. This methodology allows for the comprehensive extraction and synergistic utilization of complementary information across multimodal datasets. Extensive experiments conducted on three widely recognized HSI and LiDAR datasets demonstrate that the proposed CIGIF-Net achieves superior classification performance.}, archive = {J_ESWA}, author = {Shufang Xu and Qiyuan Xue and Zhonghao Chen and Shuyu Fei and Hongmin Gao}, doi = {10.1016/j.eswa.2025.129549}, journal = {Expert Systems with Applications}, month = {3}, pages = {129549}, shortjournal = {Expert Syst. Appl.}, title = {Complementary information-guided interactive fusion network for HSI and LiDAR data joint classification}, volume = {298}, year = {2026}, }
• (2026). GPT–empowered question–answer dataset for informative and empathetic support for korean childhood cancer survivors. ESWA, 298, 129548. (www)

  @article{ , abstract = {Despite improvements in survival rates, childhood cancer survivors in South Korea still face significant challenges in accessing the psychological and informational support they need. To address these challenges, we developed the Korean Childhood Cancer Survivor Question-Answer (KCCSQA) dataset in which contains 3876 question-answer pairs. The questions were sourced from websites, academic articles, and an online survey, where 119 childhood cancer survivors contributed 1283 questions. We used GPT-4 Turbo to generate the responses, followed by an expert evaluation by 11 specialists to ensure factual accuracy, complementarity, comprehensibility, and empathy. The overall quality of the GPT-generated responses was rated 4.98 out of 6, indicating a high level of quality. To enhance the dataset, we integrated a relational knowledge graph to mitigate hallucinations in the AI-generated answers, achieving a performance of 0.979 in hallucination detection. Additionally, a pseudo-scoring system was implemented for continuous quality assessment. The dataset’s effectiveness was evaluated through a pilot study involving 14 childhood cancer survivors, who interacted with a retrieval-based QA system using a single-turn chatbot format. The mean satisfaction rating was 4.36 on a 6-point Likert scale, and all participants expressed a willingness to use the system again.}, archive = {J_ESWA}, author = {Kyubum Hwang and Mirae Kim and Min Ah Kim and Chaerim Park and Yehwi Park and Chungyeon Lee and Jooyoung Lim and Hayoung Oh}, doi = {10.1016/j.eswa.2025.129548}, journal = {Expert Systems with Applications}, month = {3}, pages = {129548}, shortjournal = {Expert Syst. Appl.}, title = {GPT–empowered question–answer dataset for informative and empathetic support for korean childhood cancer survivors}, volume = {298}, year = {2026}, }
• (2026). A decision support system for open-pit mining optimisation with dynamic uncertainty and GPU-based parallel repair approach. ESWA, 298, 129544. (www)

  @article{ , abstract = {This paper presents an advanced Decision Support System (DSS) for long-term open-pit mine planning that integrates established optimization techniques—Large Neighborhood Search (LNS), Simulated Annealing (SA), and Dantzig-Wolfe decomposition—within a novel GPU-accelerated framework addressing geological uncertainty and computational complexity. The key methodological contributions include dynamic uncertainty modelling with time-dependent factors capturing geological confidence degradation and GPU-parallelized evaluation architecture enabling industrial-scale mine planning. Validation using 50,000 blocks across 10 geological scenarios demonstrates robust economic performance, achieving mean NPV of $1.514 billion with limited variability (standard deviation $16 million). The GPU-parallelized architecture achieves 29.6 % average computational speedup with peaks of 37 % compared to CPU implementations, enabling concurrent evaluation of 262,144 mining scenarios. Risk analysis reveals P90 Value-at-Risk of $1.488 billion, indicating strong downside protection. The system maintains profit margins exceeding 95 % across all scenarios with cumulative cash flow reaching $1.789.4 million by period 6. Narrow risk envelopes (P10-P90 spread <$60 M) demonstrate robust performance under uncertainty, providing mining companies with practical tools for risk-informed strategic decision-making.}, archive = {J_ESWA}, author = {Iman Rahimi}, doi = {10.1016/j.eswa.2025.129544}, journal = {Expert Systems with Applications}, month = {3}, pages = {129544}, shortjournal = {Expert Syst. Appl.}, title = {A decision support system for open-pit mining optimisation with dynamic uncertainty and GPU-based parallel repair approach}, volume = {298}, year = {2026}, }
• (2026). TriHID: Towards verifiable domain adaptation-based IoT intrusion detection in heterogeneous environment. ESWA, 298, 129543. (www)

  @article{ , abstract = {With the widespread deployment of IoT devices in daily applications, securing them against intrusions has become increasingly critical. Domain adaptation (DA)-based intrusion detection is a promising approach that transfers knowledge from a source domain to improve detection in a target IoT domain. However, effective DA methods must address various types of domain heterogeneity - such as differences in feature representation, intrusion distribution, and attack strategies. Existing intrusion detection datasets rarely consider these aspects, limiting their utility for evaluating heterogeneous DA approaches. To bridge this gap, we introduce TriHID , a new dataset specifically designed to capture heterogeneities from three perspectives. We evaluate four types of DA-based IoT intrusion detectors - multi-source, semi-supervised, unsupervised, and open-set on TriHID. Experimental results demonstrate that TriHID enables robust training and comprehensive evaluation of DA-based intrusion detection methods in heterogeneous IoT settings.}, archive = {J_ESWA}, author = {Jiashu Wu and Yang Wang}, doi = {10.1016/j.eswa.2025.129543}, journal = {Expert Systems with Applications}, month = {3}, pages = {129543}, shortjournal = {Expert Syst. Appl.}, title = {TriHID: Towards verifiable domain adaptation-based IoT intrusion detection in heterogeneous environment}, volume = {298}, year = {2026}, }
• (2026). Multi-objective dynamic flexible job shop scheduling using multi-head network-based deep reinforcement learning. ESWA, 298, 129542. (www)

  @article{ , abstract = {In modern manufacturing environments, job shop scheduling systems are characterized by heightened complexity and ever-changing dynamics, often involving multi-objective optimization and the need to accommodate unanticipated events like new job insertions and uncertain machine availability, underscores the necessity for effective real-time multi-objective scheduling approaches. Therefore, to tackle the multi-objective dynamic flexible job shop scheduling problem (MODFJSP) involving new job insertions, this paper introduces an online scheduling framework called multi-head deep Q network (MHDQN), designed to simultaneously minimize both total tardiness and total machine idle time. The core architecture of MHDQN framework is an innovative multi-head network agent based on Dueling deep Q network (Deuling DQN), consisting of a shared network layer and objective-specific network layers. The shared network layer extracts and transforms the input global state features layer by layer, generating a high-dimensional, semantically rich shared feature. This provides a unified input foundation for the objective-specific network layers, which are responsible for extracting the specialized information related to each objective from the shared features and calculating the corresponding Q -values, thereby enabling the parallel optimization of each objective. Six combined scheduling rules are developed to form the action set, each incorporating both job and machine selection. An improved multi-objective action selection strategy is proposed, incorporating inverse sigmoid ϵ decay and Q -value maximum absolute (max-abs) normalization to optimize decision-making. Additionally, a multi-head network training mechanism leveraging the Double deep Q network (Double DQN) architecture has been developed. Extensive computational experiments demonstrate that the MHDQN outperforms widely used traditional scheduling rules, multi-objective metaheuristic algorithms, and other reinforcement learning (RL) based scheduling methods, showing significant advantages and strong generalizability in multi-objective optimization tasks.}, archive = {J_ESWA}, author = {Kai Li and Bao Zheng and Liping Xu and Fulong Xie and Zhicheng Wang}, doi = {10.1016/j.eswa.2025.129542}, journal = {Expert Systems with Applications}, month = {3}, pages = {129542}, shortjournal = {Expert Syst. Appl.}, title = {Multi-objective dynamic flexible job shop scheduling using multi-head network-based deep reinforcement learning}, volume = {298}, year = {2026}, }
• (2026). A possibilistic programming approach in an integrated fuzzy periodic review model and clustering strategy for optimizing platelet supply chain. ESWA, 298, 129539. (www)

  @article{ , abstract = {Managing platelet supply chains poses significant challenges due to the product’s short shelf life, highly uncertain demand, and the critical nature of its medical use. Previous studies in the blood supply chain rely on fixed-order quantities and ignore collaborative inventory-sharing strategies, which can lead to either excessive waste or severe shortages. However, in many real-world situations, fixed order quantities are often insufficient to accommodate fluctuating demand, especially in healthcare systems. Moreover, existing distribution models in the literature often overlook the equitable allocation of services across hospitals, leading to disparities in access to critical healthcare resources. This study proposes a novel two-phase decision-making framework that integrates a fuzzy periodic review inventory model with a cluster-based reactive transshipment strategy to optimize platelet supply and distribution. In Phase I, a fuzzy periodic review model determines optimal order quantities under uncertain demand using a possibilistic chance-constrained programming approach. In Phase II, hospitals are clustered based on service levels, enabling equitable transshipment among facilities to reduce disparities and improve overall responsiveness. A real-world case study from Tehran province is used to evaluate the model’s effectiveness. Results show an approximate 6% reduction in total shortages and a 2% improvement in average service levels. The proposed framework offers actionable insights for healthcare managers aiming to enhance resilience, equity, and efficiency in critical medical supply chains.}, archive = {J_ESWA}, author = {Seyyed-Mahdi Hosseini-Motlagh and Mohammad Reza Ghatreh Samani and Hannaneh Kordhaghi}, doi = {10.1016/j.eswa.2025.129539}, journal = {Expert Systems with Applications}, month = {3}, pages = {129539}, shortjournal = {Expert Syst. Appl.}, title = {A possibilistic programming approach in an integrated fuzzy periodic review model and clustering strategy for optimizing platelet supply chain}, volume = {298}, year = {2026}, }
• (2026). Deep-insights guided evolutionary algorithm for optimization. ESWA, 298, 129538. (www)

  @article{ , abstract = {Evolutionary algorithms (EAs) are a class of optimization algorithms inspired by the theory of biological evolution. They solve optimization problems by emulating the processes of natural selection. EAs produce abundant data during evolution, which contains valuable information that reflects their evolutionary patterns. Effectively utilizing this information can enhance the algorithms’ effectiveness and efficiency. Deep learning excels at extracting knowledge from data. Inspired by this, we propose a novel insights-infused framework that utilizes deep neural networks to learn the evolutionary processes of EAs and extract useful synthesis insights from the evolutionary data. These synthesis insights not only guide the algorithm to evolve in a better direction on the original problems, but also improve its performance on new problems. The choice of neural networks is important. During pre-training, to reduce the inductive bias introduced by human prior knowledge, we design an MLP model to process the data. Additionally, we develop a variable-length encoding method to enable MLP networks to handle variable-length data. To verify the transfer evolution ability of synthesis insights, we devise a self-evolution strategy that fine-tunes the network using only the data generated by the algorithm itself, without introducing any external knowledge, when dealing with new problems. Experimental results demonstrate that the synthesis insights extracted from the CEC2014 dataset guide the algorithms to evolve in a better direction for the CEC2014 problems, and in addition enhance their performance on new problems like CEC2017, CEC2022 and the real-world optimization problems.}, archive = {J_ESWA}, author = {Kun Bian and Juntao Zhang and Hong Han and Jun Zhou and Yifei Sun and Shi Cheng}, doi = {10.1016/j.eswa.2025.129538}, journal = {Expert Systems with Applications}, month = {3}, pages = {129538}, shortjournal = {Expert Syst. Appl.}, title = {Deep-insights guided evolutionary algorithm for optimization}, volume = {298}, year = {2026}, }
• (2026). MambaGen: Efficient visual representation learning for automatic radiology report generation. ESWA, 298, 129537. (www)

  @article{ , abstract = {Radiology report generation focuses on producing comprehensive and clinically precise medical reports based on radiographic images, thereby improving medical efficiency and alleviating the burden on radiologists. Although existing deep learning methods have demonstrated superior performance, they are constrained by the local receptive field of convolutional neural networks and are inadequate for modeling long-range dependencies, making it challenging to detect critical lesion features in medical images. Recently, State Space Models (SSMs), particularly Mamba, have shown great potential in modeling long-range dependencies with linear computational complexity. Inspired by this, we propose MambaGen, the enhanced Mamba model specifically designed for radiology report generation tasks. Specifically, we design a Mamba-Visual Recalibration Module (MVRM), which utilizes a two-stage training strategy to effectively capture the efficient visual representation of medical images. This first stage combines convolutional layers with SSMs to model long-sequence dependencies and learn multi-level visual feature information. This second stage introduces local convolution and a channel attention mechanism to further recalibrate the local feature and mitigate channel redundancy. Comprehensive experiments on widely available datasets, such as IU X-Ray and MIMIC-CXR, demonstrate our model’s superior performance compared to existing methods, particularly with an improvement of 2.7 % on the BLEU-4 metric. The code is available at https://github.com/Eleanorhxd/MambaGen.git .}, archive = {J_ESWA}, author = {Xiaodi Hou and Xiaobo Li and Simiao Wang and Mingyu Lu and Hongfei Lin and Yijia Zhang}, doi = {10.1016/j.eswa.2025.129537}, journal = {Expert Systems with Applications}, month = {3}, pages = {129537}, shortjournal = {Expert Syst. Appl.}, title = {MambaGen: Efficient visual representation learning for automatic radiology report generation}, volume = {298}, year = {2026}, }
• (2026). Multistep diesel vehicle emissions forecasting with an efficient transformer enhanced by temporal-frequency fusion and covariate interaction. ESWA, 298, 129532. (www)

  @article{ , abstract = {Controlling nitrogen oxide (NO x ) emissions from diesel vehicles is a critical environmental challenge. Advanced SCR strategies depend on accurate multistep forecasting of NO x and ammonia (NH 3 ), but existing models often struggle with error accumulation and the non-stationary dynamics of emissions data. In this work, we first establish a new benchmark for this task, confirming that handling data non-stationarity is essential for robust prediction. Building on this insight, we propose FiTformer, a novel Transformer architecture that adopts an encoder-only framework to jointly forecast both NO x and NH 3 concentrations, where we introduce Intra-series Temporal-Frequency Fusion mechanism to capture intrinsic emissions dynamics and Inter-series Covariate Interaction mechanism to model external influences. Validated on real-world engine data, FiTformer consistently outperforms baseline models across all evaluated prediction horizons, with up to 44.1 % MAE and 36.9 % SMAPE reductions in 24-step NO x prediction and similarly strong gains for NH 3 prediction, compared to the state-of-the-art baseline TimesNet. Its high computational efficiency (0.30G MACs and 8.3 ms/iter) along with robust generalization and high resilience to data imperfections, underscores its suitability for real-time embedded SCR control, enabling more effective strategies for NO x reduction and NH 3 slip minimization.}, archive = {J_ESWA}, author = {Yuhan Luo and Yujun Zhang and Ying He and Kun You and Wei Huang and Wenqing Liu and Hao Xie}, doi = {10.1016/j.eswa.2025.129532}, journal = {Expert Systems with Applications}, month = {3}, pages = {129532}, shortjournal = {Expert Syst. Appl.}, title = {Multistep diesel vehicle emissions forecasting with an efficient transformer enhanced by temporal-frequency fusion and covariate interaction}, volume = {298}, year = {2026}, }
• (2026). An EEG-based dual-stream spatial-spectral-temporal large model for self-limited epilepsy with centrotemporal spikes. ESWA, 298, 129530. (www)

  @article{ , abstract = {Self-limited epilepsy with centrotemporal spikes (SeLECTS) is the most common form of focal epilepsy in childhood, accounting for 20–25 % of all childhood epilepsy cases and may be associated with cognitive dysfunction and behavioral issues. Accurate detection and assessment of epileptic discharges in EEG signals, particularly the spike-wave index (SWI), are crucial for timely intervention and treatment. Manual analysis of EEG data is labor-intensive and prone to errors, underscoring the need for automated methods. In the present study, we propose a novel D ual-Str e am Sp a tial- S pectral- T emporal L arge model (DeaSTL) that leverages a large-scale EEG architecture to effectively capture the multidimensional characteristics of EEG signals associated with SeLECTS syndrome. Our model integrates multi-view temporal representations and spatial-spectral representations through a dual-stream approach, enhancing the learning of complex patterns in EEG data. We introduce the S JTU Se L ECTS E EG D ataset (SLED), a comprehensive EEG dataset from 212 patients diagnosed with SeLECTS, including annotations for abnormal discharge detection, wake-sleep period classification, and SWI estimation. Addressing the previously unexplored problem of SWI prediction, we provide a novel method for quantifying the severity of epileptic discharges during sleep. Extensive experiments demonstrate that our DeaSTL model significantly outperforms several state-of-the-art methods across multiple tasks, showcasing its potential for clinical application in assisting diagnosis and treatment planning.}, archive = {J_ESWA}, author = {Lin Zhang and Yun Ren and Fang Yuan and Xuqin Chen and Shikui Tu and Lei Xu}, doi = {10.1016/j.eswa.2025.129530}, journal = {Expert Systems with Applications}, month = {3}, pages = {129530}, shortjournal = {Expert Syst. Appl.}, title = {An EEG-based dual-stream spatial-spectral-temporal large model for self-limited epilepsy with centrotemporal spikes}, volume = {298}, year = {2026}, }
• (2026). Dynamic learning of sample ambiguity-driven sample weighting for medical image classification. ESWA, 298, 129527. (www)

  @article{ , abstract = {Deep neural networks have delivered impressive results in medical image classification tasks. However, their performance is still challenging in medical scenarios with limited data, where training set biases such as label noise or class imbalance impede model learning. Dynamic learning based sample weighting achieves adaptive adjustment of sample importance through learnable weight functions and shows great potential in improving model robustness. Nevertheless, existing methods directly employ model states such as loss value or training accuracy to evaluate sample importance, ignoring the role of ambiguous samples in model optimization. This limitation hinders the performance of dynamic learning based sample weighting in medical image classification. In this paper, we propose a new sample weighting approach based on sample ambiguity and dynamic learning for improving medical image classification, named DLSA-SW. We introduce a dual-space sample ambiguity method by evaluating the category proximity in the feature space and the prediction confidence in the label space. Subsequently, to dynamically calculate sample weights according to sample ambiguity, a learnable sample weighting network is developed to adaptively adjust the weights during training to guide the task model. DLSA-SW performs alternate optimization to enable mutual adaptation of the sample weighting network and the task network. We evaluate the effectiveness of our approach on three medical image classification benchmarks: PatchCamelyon for lymph node histopathology classification, ISIC 2020 for skin lesion classification, and MTC for medullary thyroid cancer classification. DLSA-SW outperforms existing state-of-the-art sample weighting methods on all three datasets and yields substantial improvements over methods without sample weighting. These results demonstrate the robustness and practical applicability of our approach in clinical diagnostic tasks.}, archive = {J_ESWA}, author = {Guanxiu Yi and Xiabi Liu and Ling Ma and Mengqiao Han and Lijuan Niu}, doi = {10.1016/j.eswa.2025.129527}, journal = {Expert Systems with Applications}, month = {3}, pages = {129527}, shortjournal = {Expert Syst. Appl.}, title = {Dynamic learning of sample ambiguity-driven sample weighting for medical image classification}, volume = {298}, year = {2026}, }
• (2026). LST-rPPG: A long-range spatio-temporal model for high-accuracy heart rate variability measurement. ESWA, 298, 129526. (www)

  @article{ , abstract = {Remote photoplethysmography (rPPG) is a non-contact manner of measuring heart rate variability (HRV) by deriving blood volume pulse (BVP) signals from facial videos. The performance of rPPG-based HRV measurement is challenging due to short-range noises (e.g., head movements) suppression and sufficient-duration BVP signal generation. Recent Transformer-based rPPG methods have shown advantages of global spatio-temporal feature modeling in eliminating noise and recovering high-quality BVP signals. However, these methods often face significant computational and memory constraints, limiting duration scalability of the generated BVP signals that might decrease HRV measurement performance. To address the above issue, this paper proposes a duration-scalable Transformer-based rPPG method, capable of global Long-range Spatio-Temporal modelling (LST), termed LST-rPPG, to generate high-quality BVP signals with a much longer duration. On the one hand, by employing spatial and temporal encoders, the original image-based rPPG issue is converted to a time-series problem. Besides, a sparse computation mechanism is integrated into the temporal encoder. This combination allows LST-rPPG to recover flexible-duration BVP signals, supporting continuous modeling of segments beyond 30 s with low computation and memory overhead. On the other hand, a dynamic loss function with stringent temporal constraints is designed to guarantee the quality of the generated BVP signals. Comprehensive experiments are performed on two public datasets, PURE and UBFC-RPPG, and the results demonstrate the feasibility of LST-rPPG for generating high-quality BVP signals with a much longer duration while requiring substantially fewer computational resources. Besides, LST-rPPG achieves at least the second-best results during all experiments.}, archive = {J_ESWA}, author = {Jiajie Li and Juan Cheng and Rencheng Song and Yu Liu}, doi = {10.1016/j.eswa.2025.129526}, journal = {Expert Systems with Applications}, month = {3}, pages = {129526}, shortjournal = {Expert Syst. Appl.}, title = {LST-rPPG: A long-range spatio-temporal model for high-accuracy heart rate variability measurement}, volume = {298}, year = {2026}, }
• (2026). A novel scheme integrating graph-based analysis and opposition-based learning for S-box optimization by population-based metaheuristics. ESWA, 298, 129524. (www)

  @article{ , abstract = {The non-linearity of symmetric block cryptographic algorithms, which is crucial for resisting linear and differential cryptanalysis, depends on the design of substitution boxes. This work proposes a novel scheme, named GOM, combining opposition-based learning with graph-based representation to be integrated into three population-based metaheuristics. This scheme is applied to address the 8 × 8 S-box design problem. The GOM-enhanced metaheuristics improve population diversity and convergence, achieving a non-linearity of 112. Hardware simulations indicate that GOM-based designs require similar resources to AES, while side-channel evaluations confirm their resilience against power analysis attacks. Scalability is supported by successful results on 10 × 10 and 12 × 12 S-boxes. The design of GOM, leveraging generalizable components such as opposition-based learning and graph-based representations, suggests its potential applicability to other population-based metaheuristics and optimization problems.}, archive = {J_ESWA}, author = {Francisco González and Ricardo Soto and José M. Lanza-Gutiérrez and Broderick Crawford}, doi = {10.1016/j.eswa.2025.129524}, journal = {Expert Systems with Applications}, month = {3}, pages = {129524}, shortjournal = {Expert Syst. Appl.}, title = {A novel scheme integrating graph-based analysis and opposition-based learning for S-box optimization by population-based metaheuristics}, volume = {298}, year = {2026}, }
• (2026). Functional consistency of LLM code embeddings: A self-evolving data synthesis framework for benchmarking. ESWA, 298, 129523. (www)

  @article{ , abstract = {Embedding models have demonstrated strong performance in tasks like clustering, retrieval, and feature extraction while offering computational advantages over generative models and cross-encoders. Benchmarks such as MTEB have shown that text embeddings from large language models (LLMs) capture rich semantic information, but their ability to reflect code-level functional semantics remains unclear. Existing studies largely focus on code clone detection, which emphasizes syntactic similarity and overlooks functional understanding. In this paper, we focus on the functional consistency of LLM code embeddings, which determines if two code snippets perform the same function regardless of syntactic differences. We propose a novel data synthesis framework called Functionality-Oriented Code Self-Evolution to construct diverse and challenging benchmarks. Specifically, we define code examples across four semantic and syntactic categories and find that existing datasets predominantly capture syntactic properties. Our framework generates four unique variations from a single code instance, providing a broader spectrum of code examples that better reflect functional differences. Extensive experiments on three downstream tasks-code clone detection, code functional consistency identification, and code retrieval-demonstrate that embedding models significantly improve their performance when trained on our evolved datasets. These results highlight the effectiveness and generalization of our data synthesis framework, advancing the functional understanding of code.}, archive = {J_ESWA}, author = {Zhuohao Li and Wenqing Chen and Jianxing Yu and Zhichao Lu}, doi = {10.1016/j.eswa.2025.129523}, journal = {Expert Systems with Applications}, month = {3}, pages = {129523}, shortjournal = {Expert Syst. Appl.}, title = {Functional consistency of LLM code embeddings: A self-evolving data synthesis framework for benchmarking}, volume = {298}, year = {2026}, }
• (2026). Semisupervised graph U-net with G-ConvLSTM for hyperspectral image classification. ESWA, 298, 129522. (www)

  @article{ , abstract = {Graph convolutional networks (GCNs) have demonstrated the effectiveness for hyperspectral image (HSI) classification, but still face challenges, such as insufficient exploitation of data structure information, limited labeled samples, and high susceptibility to noise and outliers. To address these issues, a semisupervised graph U-Net with graph convolutional long short-term memory is proposed for HSI classification, abbreviated as SSGU-Net. Specifically, we design a novel graph convolutional long short-term memory feature extractor to learn discriminative spatial-spectral joint features by simultaneously modeling the correlations in the spatial and spectral domains. Then, we develop a semisupervised graph U-Net with mutually inverse operation of the graph pooling and the graph unpooling modules which uses both labeled samples and unlabeled samples to train a well-parameterized network for HSI classification. In particular, to suppress the effects of noise and outliers, the graph pooling module is designed to selectively retain discriminative samples and fully learn the optimal correlation between these retained samples. Meanwhile, the graph unpooling module employs the local spatial context to reconstruct the reduced samples, thus restoring the pooled data to its original scale for classification task. Extensive experiments show the effectiveness of the proposed method, achieving overall accuracy gains of 5.22 %, 1.58 %, and 1.57 % over the state-of-the-art competitors on the Indian Pines, University of Pavia, and Houston datasets, respectively.}, archive = {J_ESWA}, author = {Jin-Yu Yang and Heng-Chao Li and Xin-Ru Feng and Feng Gao and Qian Du and Antonio Plaza}, doi = {10.1016/j.eswa.2025.129522}, journal = {Expert Systems with Applications}, month = {3}, pages = {129522}, shortjournal = {Expert Syst. Appl.}, title = {Semisupervised graph U-net with G-ConvLSTM for hyperspectral image classification}, volume = {298}, year = {2026}, }
• (2026). Chat with one voice: Mitigating semantic disparity across languages for multilingual open-domain dialogue response generation systems. ESWA, 298, 129521. (www)

  @article{ , abstract = {With the development of the Internet, building a multilingual dialogue generation system to attract more users while reducing costs in the global market has become increasingly important. However, current end-to-end multilingual approaches often face semantic disparity issues across languages. When given parallel queries with the same semantics but in different languages, the generated responses may vary in meaning across languages, which may greatly affect the stability and reliability of multilingual systems in different language scenarios. We attribute this issue to open-domain/model uncertainty and language differences. To mitigate this issue, we first propose a novel Anchor-based Semantic Constraint (ASC) designed to reduce semantic disparity across languages for Encoder-Decoder Transformers. ASC employs language-independent anchor signal to guide the behaviors in both the encoder and decoder, thereby reducing uncertainty. Additionally, ASC incorporates a two-stage tuning process to further minimize the impact of language differences by ensuring the encoder remains language-independent. Extensive experiments and in-depth analyses conducted on XDailyDialog demonstrate that ASC can effectively mitigate semantic disparity across languages and will not compromise dialogue response quality like the previous related baselines.}, archive = {J_ESWA}, author = {Sixing Wu and Jiahao Chen and Jiong Yu and Wei Zhou}, doi = {10.1016/j.eswa.2025.129521}, journal = {Expert Systems with Applications}, month = {3}, pages = {129521}, shortjournal = {Expert Syst. Appl.}, title = {Chat with one voice: Mitigating semantic disparity across languages for multilingual open-domain dialogue response generation systems}, volume = {298}, year = {2026}, }
• (2026). Towards CPU performance prediction: New challenge benchmark dataset and novel approach. ESWA, 298, 129520. (www)

  @article{ , abstract = {CPU performance prediction based on hardware characteristics is crucial for system design and resource management. However, this field faces two major challenges. First, collecting real-world data is challenging due to the diversity of CPU products and the specialized nature of hardware characteristics. This field lacks a standard dataset with unified hardware characteristics, wide data coverage, and comprehensive benchmarks. Second, existing methods based on hardware simulation models or machine learning exhibit notable shortcomings, such as lengthy simulation test cycles and low prediction accuracy. To bridge these gaps, we first collect, preprocess, and standardize historical data from the 4th Generation Intel® Xeon® Scalable Processors across multiple benchmark suites to create a new dataset, named PerfCastDB. Subsequently, we design a deep learning based model called Nova CPU Performance Predictor (NCPP) as the baseline for this new dataset. The NCPP network is designed based on group attention mechanism. It effectively quantifies the implicit relationships between hardware characteristics within and across groups and comprehensively models the impact of various hardware characteristics on CPU performance prediction. We conduct comparative experiments using the proposed PerfCastDB dataset. Compared to existing approaches, NCPP achieves superior evaluation results, demonstrating its effectiveness. Furthermore, we have open-sourced part of the dataset and the NCPP network code to facilitate subsequent research. The resources can be accessed at https://github.com/xiaoman-liu/NCPP .}, archive = {J_ESWA}, author = {Xiaoman Liu}, doi = {10.1016/j.eswa.2025.129520}, journal = {Expert Systems with Applications}, month = {3}, pages = {129520}, shortjournal = {Expert Syst. Appl.}, title = {Towards CPU performance prediction: New challenge benchmark dataset and novel approach}, volume = {298}, year = {2026}, }
• (2026). Glyph graph isomorphism network for structure recognition of oracle bone inscription. ESWA, 298, 129519. (www)

  @article{ , abstract = {Structure recognition of oracle bone inscription glyphs plays an important role in studying the evolutionary process of oracle bone inscriptions and the history of the Shang Dynasty. Currently, most methods have decomposed oracle bone inscription glyphs into multilevel features, which are used to recognize hierarchical feature fusion. This strategy cannot recognize the primitive internal structures of keypoints, strokes, and components. Moreover, mainstream graph neural networks cannot fully utilize the rich structural information of oracle bone inscription glyphs, resulting in their inability to meet the needs of structure recognition. So we have developed a graph structure recognition method to implement structure recognition of oracle bone inscription glyphs. A graph extraction method is given to get the structure of oracle bone inscription glyphs; each graph structure’s representation vector can be learned by a glyph graph isomorphism network, which is developed to recognize the graph of oracle bone inscription glyphs to enhance the discriminability representation of the structure. Our model has achieved advanced results across structure recognition experiments in the HWOBC dataset and the Oracle-50K dataset.}, archive = {J_ESWA}, author = {Zhan Zhang and Hanbin Liu and Xingkun Zhang and Yiyuan Wang and Feng Gao and An Guo and Han Zhang and Qingju Jiao and Bang Li and Yongge Liu}, doi = {10.1016/j.eswa.2025.129519}, journal = {Expert Systems with Applications}, month = {3}, pages = {129519}, shortjournal = {Expert Syst. Appl.}, title = {Glyph graph isomorphism network for structure recognition of oracle bone inscription}, volume = {298}, year = {2026}, }
• (2026). An integrated decision-making framework to evaluate the route alternatives in overweight/oversize transportation. ESWA, 298, 129516. (www)

  @article{ , abstract = {Overweight and oversized transport (O&OT) has become one of the most critical elements of project logistics, driven by advancements in transportation and lifting technologies that now allow high-volume loads to be moved across long distances. This type of transportation operation, also called abnormal transportation, is greatly affected by technical factors such as the weight and geometry of the load, road surface, axle load limitations, slope, and ground strength, as well as external variables such as weather conditions, traffic density, and legal regulations. In planning and operational processes, Decision-Makers (DMs) and practitioners who plan and execute operations without adequately considering these factors and variables can lead to delays in operations, serious risks, and loss of productivity. This research proposes a flexible decision support model that integrates Step-wise Weight Assessment Ratio Analysis (SWARA) and Logarithmic Percentage Change-driven Objective Weighting (LOPCOW), and a ranking technique; i.e., Mixed Aggregation by Comprehensive Normalization Technique (MACONT) techniques to address the decision problems related to route selection, one of the most critical problems in transporting heavy and bulky loads, and to produce reasonable solutions. The proposed model significantly reduces information losses by processing subjective and objective information and integrating subjective (SWARA) and objective (LOPCOW) methods. Unlike traditional ranking approaches, the MACONT method combines three different normalization techniques to determine the ranking performance of alternatives. In this way, it provides more reliable and accurate results by reducing the deviations of the results provided by the single normalization technique. In addition, it shows each alternative’s good and bad performance compared to the others and is more convincing about the results obtained. According to the results obtained by applying the proposed model, fuel consumption (0.096) is determined as the most effective and critical factor in selecting the route on which heavy and bulky loads will be transported. In this context, choosing routes that allow lower fuel consumption can contribute to reducing carbon emissions and external costs arising from transportation. The extensive robustness and validation check to test the proposed model prove that the proposed model is a reliable, robust, and practical decision-making tool for making reasonable and rational decisions in O&OT.}, archive = {J_ESWA}, author = {Ömer Faruk Görçün and Pradip Kundu and Hande Küçükönder and Gürkan Doğan and Erfan Babaee Tirkolaee}, doi = {10.1016/j.eswa.2025.129516}, journal = {Expert Systems with Applications}, month = {3}, pages = {129516}, shortjournal = {Expert Syst. Appl.}, title = {An integrated decision-making framework to evaluate the route alternatives in overweight/oversize transportation}, volume = {298}, year = {2026}, }
• (2026). EvoMapX: An explainable framework for metaheuristic optimization algorithms. ESWA, 298, 129514. (www)

  @article{ , abstract = {Population-based optimization algorithms (POAs) are widely adopted solutions for NP-hard and complex high-dimensional optimization problems. However, their internal dynamics often remain opaque, limiting trust and insight into how solutions evolve. This paper introduces EvoMapX, a novel explainable framework designed to interpret the internal dynamics of population-based optimization algorithms. EvoMapX includes three interpretable structures to visualize evolutionary optimization dynamics: the Operator Attribution Matrix (OAM) quantifies the contribution of specific operators over iterations; the Population Evolution Graph (PEG) traces the ancestry and transformation of candidate solutions; and the Convergence Driver Score (CDS) identifies which operators drive convergence, helping interpret why the algorithm improved. EvoMapX was evaluated across four POAs on the CEC 2021 test suite in order to demonstrate how it reveals meaningful textual and graphical insights into algorithm behavior. EvoMapX paves the way for interpretable metaheuristic optimization. The source code of EvoMapX is available at https://www.github.com/Bilal20252025/EvoMapX}, archive = {J_ESWA}, author = {Bilal H. Abed-Alguni}, doi = {10.1016/j.eswa.2025.129514}, journal = {Expert Systems with Applications}, month = {3}, pages = {129514}, shortjournal = {Expert Syst. Appl.}, title = {EvoMapX: An explainable framework for metaheuristic optimization algorithms}, volume = {298}, year = {2026}, }
• (2026). BDOS: An orthogonal sum based on belief of permutation events and element distances in random permutation set. ESWA, 298, 129513. (www)

  @article{ , abstract = {The random permutation set (RPS) extends Dempster-Shafer evidence theory by incorporating event order information, providing a powerful framework for modeling uncertainty. However, existing orthogonal sum methods within the RPS framework may encounter loss of order information and counterintuitive belief distribution during permutation event fusion. To address these two issues, this paper proposes a new method termed belief-distance-based orthogonal sum (BDOS). BDOS operates through three core mechanisms: order-information preservation via mathematical constructs like order-space and inverse mapping; belief-value weighting that prioritizes events with high belief mass for rational outcomes; and element-distance weighting that incorporates dissimilarity among permutations to improve ordinal accuracy. Numerical examples validate the effectiveness of BDOS in permutation event fusion, with comparative results demonstrating its advantages in order retention and belief distribution. Furthermore, BDOS is applied to threat assessment, illustrating its rationality and effectiveness in handling uncertainty and threat ranking.}, archive = {J_ESWA}, author = {Xiaoyan Su and Xu Chen and Hong Qian and Cheng Jiang}, doi = {10.1016/j.eswa.2025.129513}, journal = {Expert Systems with Applications}, month = {3}, pages = {129513}, shortjournal = {Expert Syst. Appl.}, title = {BDOS: An orthogonal sum based on belief of permutation events and element distances in random permutation set}, volume = {298}, year = {2026}, }
• (2026). Fine-grained similarity retrieval of lesion areas in lung CT images based on visual similarity matching of image blocks. ESWA, 298, 129510. (www)

  @article{ , abstract = {Background and Objective With the rapid growth in the number of medical images, the need for content- based medical image retrieval (CBMIR) in clinical aid diagnosis is becoming increasingly important. Most current content-based CT image similarity retrieval methods use the entire CT image, ignoring the fact that the localized lesion region is the main target of similarity retrieval; Methods To address this issue, the paper proposes a fine-grained similarity retrieval method for lung CT images based on image block( IB ) similarity matching, taking lung CT images as an example. In this method, two enabling techniques are introduced: 1) a hybrid Convolution and Vision Transformer Model(CVTM) that effectively captures both local texture and global context features of lesion regions; 2) the iDS high-dimensional index designed to accelerate retrieval among IB ; Results With the aid of these techniques, fine-grained similarity retrieval optimization of lung CT images can be achieved, which facilitates more accurate lesion-level comparison and supports clinical decision-making; Conclusions Extensive experiments are conducted to indicate that the proposed fine-grained similarity retrieval method achieves excellent performance, with a mAP of 91.33%. Meanwhile, the retrieval efficiency of the iDS high-dimensional index is about 150% higher than that of sequential retrieval, especially when the retrieval radius is large and the database size is substantial.}, archive = {J_ESWA}, author = {Yi Zhuang and Jiayu Zhang and Yujia Ge and Nan Jiang}, doi = {10.1016/j.eswa.2025.129510}, journal = {Expert Systems with Applications}, month = {3}, pages = {129510}, shortjournal = {Expert Syst. Appl.}, title = {Fine-grained similarity retrieval of lesion areas in lung CT images based on visual similarity matching of image blocks}, volume = {298}, year = {2026}, }
• (2026). Traffic prediction using an active causality recurrent graph convolutional network. ESWA, 298, 129506. (www)

  @article{ , abstract = {The quality of our daily lives is significantly influenced by traffic conditions, highlighting the importance of incorporating complex spatiotemporal dependencies in interconnected traffic data for effective prediction. Although recent advancements have demonstrated prediction accuracy using graph convolutional networks, their depends heavily on the accuracy of the graph structures that represent the spatial relationships within the traffic network. To address this challenge, we introduce a novel approach to traffic prediction, the Active Causal Recurrent Graph Convolution Network (ACRGCN), as shown in Fig. 2. ACRGCN offers a new framework that effectively integrates a causal-embedded approach for traffic prediction, leveraging both structural and feature information from correlated traffic time series. Additionally, it incorporates a time-varying dynamic Bayesian network to capture the intricate spatiotemporal topology of traffic data. The model extracts spatiotemporal dependencies from traffic signals using the Active Causality Graph Recurrent Module (ACGRM), while efficiently modeling nonlinear traffic propagation patterns. Furthermore, ACRGCN employs a deep learning-based module that functions as a hyper-network, progressively generating dynamic causal graphs. Finally, extensive experiments on multiple real-world traffic graph datasets validate ACRGCN, and the results demonstrate its superiority over state-of-the-art method}, archive = {J_ESWA}, author = {Jinde Zhu and Junhao Yuan and Fulan Ye and Trong-The Nguyen and Ruoxi Wang and Wu Zeng and Chien-Chun Liu}, doi = {10.1016/j.eswa.2025.129506}, journal = {Expert Systems with Applications}, month = {3}, pages = {129506}, shortjournal = {Expert Syst. Appl.}, title = {Traffic prediction using an active causality recurrent graph convolutional network}, volume = {298}, year = {2026}, }
• (2026). Shielding federated learning: Mitigating label transferability against poisoning attacks in cloud-edge-client system. ESWA, 298, 129502. (www)

  @article{ , abstract = {Hierarchical computing architecture of cloud-edge-client (CEC) formed with the combination of cloud computing and edge computing can provide processing, storage and low-latency services close to end devices. To protect data privacy, federated learning (FL), as a novel intelligent edge computing framework with localized training mechanisms, has been integrated into edge computing to form a system called CEC-FL and is widely studied. However, they are susceptible to potential poisoning attacks. Existing poisoning attack methods are mostly explored by performing malicious operations on training samples or labels directly and implementing corresponding defense strategies: they are designed to ignore the label transferability and diverse attack environments and are not work against stealthy security threats, mainly because they do not take into account the inherent vulnerabilities of the attack environment. Yet few general defense schemes have been developed. In response to the above vulnerabilities, in this work, we explore a B arycenter Po isoning method with L abel T ransferability (BPoLT) initiated by malicious attackers, resulting in a dynamic attack capability on the CEC-FL system. To address poisoning attacks, we provide a two-phase defense algorithm Res isting L abel T ransferability Pois oning called ResLT-Pois to distinguish malicious attackers from benign participants. Extensive experimental results demonstrate that our scheme is feasible and effective in dealing with the vulnerability of the CEC-FL system.}, archive = {J_ESWA}, author = {Yaru Zhao and Yihao Cao and Jianbiao Zhang and Zhaoqian Zhang and Weiru Wang}, doi = {10.1016/j.eswa.2025.129502}, journal = {Expert Systems with Applications}, month = {3}, pages = {129502}, shortjournal = {Expert Syst. Appl.}, title = {Shielding federated learning: Mitigating label transferability against poisoning attacks in cloud-edge-client system}, volume = {298}, year = {2026}, }
• (2026). SAMForecast: A hybrid model of self-attention and mamba with adaptive wavelet transform for time series forecasting. ESWA, 298, 129498. (www)

  @article{ , abstract = {Time series forecasting faces significant challenges due to the variability and complexity of real-world data. Traditional methods often require manual adjustments of wavelet transform parameters, which are labor-intensive and prone to over-fitting or inadequate feature extraction. To address these limitations, this study proposes SAMForecast, a novel hybrid model that integrates Adaptive Wavelet Transform, self-attention mechanisms, and the selective state space model. We introduce an Adaptive Wavelet Block that dynamically adjusts decomposition levels and basis functions using a Mixture of Experts network and lifting scheme, eliminating the need for manual parameter tuning. Furthermore, the model deeply integrates the attention mechanism of the Transformer architecture, leveraging its advantages in capturing complex dependencies to identify correlations between time series data. By combining self-attention with Mamba, SAMForecast effectively captures both global dependencies and local key features in time series, enhancing robustness against noise and redundant information. SAMForecast demonstrates promising performance in multivariate time series forecasting tasks, showcasing an average 2 % performance improvement compared to existing models across datasets in energy, transportation, and other fields. The code is available at https://github.com/Kiki-V/SAMForecast-main .}, archive = {J_ESWA}, author = {Dunlu Peng and Qiqi Lin}, doi = {10.1016/j.eswa.2025.129498}, journal = {Expert Systems with Applications}, month = {3}, pages = {129498}, shortjournal = {Expert Syst. Appl.}, title = {SAMForecast: A hybrid model of self-attention and mamba with adaptive wavelet transform for time series forecasting}, volume = {298}, year = {2026}, }
• (2026). An adaptive hypergraph-based convolution network with dual spatiotemporal attention for PM2.5 forecasting. ESWA, 298, 129497. (www)

  @article{ , abstract = {Accurate PM2.5 prediction is crucial for effective environmental management and public health protection, yet current models show limited dynamic adaptability to complicated air pollution scenarios. Robust models are essential to support timely interventions in response to sudden pollution events or rapidly changing air quality patterns. However, existing models predominantly rely on predefined graph structures and pairwise spatial relationships, limiting their ability to capture the complex and dynamic interactions inherent in PM2.5 pollution. Furthermore, such models often assume equal contributions from neighboring nodes, neglecting heterogeneity and compromising predictive accuracy. To address these limitations, we propose an Adaptive Hypergraph-based Convolution Network with a Dual Spatiotemporal Attention mechanism (AHCN-DA) for PM2.5 forecasting. This framework leverages representation learning and hypergraph structures to capture and integrate pairwise as well as higher-order spatial interactions, producing richer spatial-feature representations. The dual spatiotemporal attention mechanism dynamically assigns time-varying weights to neighboring nodes based on their relevance to target nodes, effectively mitigating the impact of irrelevant inputs. Additionally, AHCN-DA integrates a dilated convolution network with multi-scale kernels to capture temporal patterns effectively across varying scales. Extensive experiments on the 2023 China National Air Quality Dataset show significant improvements in predictive accuracy, particularly in enhancing the proportion of high-precision monitoring stations, with an R 2 of 0.9224, outperforming baseline models. Our findings underscore the effectiveness of AHCN-DA in enhancing prediction accuracy under complex pollution response patterns, contributing to more informed decision-making in environmental management.}, archive = {J_ESWA}, author = {Haipeng Gao and Chonghui Qian and Yang Su and Wei Zhang and Hengjun Huang}, doi = {10.1016/j.eswa.2025.129497}, journal = {Expert Systems with Applications}, month = {3}, pages = {129497}, shortjournal = {Expert Syst. Appl.}, title = {An adaptive hypergraph-based convolution network with dual spatiotemporal attention for PM2.5 forecasting}, volume = {298}, year = {2026}, }
• (2026). Solving equilibrium for adversarial team games utilizing fictitious team play with refined team plans. ESWA, 298, 129496. (www)

  @article{ , abstract = {Adversarial team games represent scenarios where cooperation and competition coexist and hold numerous applications in the real world. These scenarios are particularly challenging due to asymmetric information among team members and limited communication capabilities. Fictitious team-play extend self-play algorithms to these scenarios, offering a novel approach to obtain equilibrium. However, it depends on normal-form team plans, which expand exponentially with game size, significantly constraining their applicability in large games. To overcome the challenge of computing equilibrium in large scale imperfect information team games, we propose a team self-play algorithm that utilizes refined team plans. Specifically, we pre-solve the equilibrium in a perfect recall environment to extract essential team plans from the original strategy space. To adapt these plans to an imperfect recall environment, we construct an auxiliary game with transformed ex ante coordinated information based on the original game and then solve equilibrium in auxiliary game to derive equilibrium for the original game. The experiments demonstrate the effectiveness of our team self-play algorithm in eight different Kuhn poker scenarios. Compared to existing team self-play algorithms, our method efficiently handles large games and exhibits superior convergence compared to reinforcement learning based algorithms. Additionally, our experiments offer valuable insights and guidance on adapting equilibrium strategies from perfect recall environments to those with imperfect recall.}, archive = {J_ESWA}, author = {Jinheng Xiao and Chen Qiu and Yingying Xu and Jiajia Zhang and Shuhan Qi and Xuan Wang}, doi = {10.1016/j.eswa.2025.129496}, journal = {Expert Systems with Applications}, month = {3}, pages = {129496}, shortjournal = {Expert Syst. Appl.}, title = {Solving equilibrium for adversarial team games utilizing fictitious team play with refined team plans}, volume = {298}, year = {2026}, }
• (2026). Proactive-BiLSTM: Glaucoma detection using FUNDUS and OCT retinal images. ESWA, 298, 129490. (www)

  @article{ , abstract = {Glaucoma detection identifies the early signs of eye conditions that can lead to vision loss by analyzing the retinal images to detect abnormalities, such as increased intraocular pressure, changes in the optic nerve head, or structural alterations in the retina. The challenges faced by the existing models include the difficulty in detecting subtle features, variability in image quality, and complex patterns that may resemble normal variations. Moreover, the traditional models struggle to adapt to the evolving patient data, capture long-term dependencies, and often suffer from lower accuracy. Hence, this research proposes the Proactive Hybridized Bidirectional Long Short-Term Memory (BiLSTM) model for Glaucoma detection. The proactive hybridized BiLSTM model is designed to enhance the detection of glaucoma by processing the retinal images. The proactive hybridized BiLSTM model enables the model to capture complex temporal dependencies and relationships within the data, which are crucial for identifying subtle patterns indicative of glaucoma, for which multifaceted feature extraction is employed. Moreover, the Proactive Hybridized BiLSTM model adapts to dynamic changes in the data to learn and predict glaucoma-related features, ultimately improving detection performance over time. The proposed Proactive hybridized BiLSTM model attains higher accuracy, sensitivity, and specificity of 96.65%, 96.51%, and 96.79% using the OCT and FUNDUS image dataset.}, archive = {J_ESWA}, author = {M.Kiran Mayee and M.Humera Khanam and Shaik Lathifa Tabasum}, doi = {10.1016/j.eswa.2025.129490}, journal = {Expert Systems with Applications}, month = {3}, pages = {129490}, shortjournal = {Expert Syst. Appl.}, title = {Proactive-BiLSTM: Glaucoma detection using FUNDUS and OCT retinal images}, volume = {298}, year = {2026}, }
• (2026). DBCD: Deep balanced community detection via consensus-guided joint optimization in attributed networks. ESWA, 298, 129487. (www)

  @article{ , abstract = {Current deep learning methods for community detection in attributed networks face a critical limitation: they often fail to identify communities that are both structurally cohesive and semantically similar, thereby falling short of the balance typically observed in human-labeled partitions. This shortcoming stems from the absence of explicit mechanisms to jointly optimize these two objectives. In this paper, this challenge is addressed by proposing Deep Balanced Community Detection (DBCD), a novel unsupervised framework for community detection that balances topology and semantics. DBCD first constructs a powerful topology-semantic clustering consensus by integrating insights from both structural and attribute spaces. This consensus then steers a Graph Neural Network to simultaneously maximize global neural modularity and local cross-view consistency, while adaptively determining the number of communities. Extensive experiments reveal a striking result: DBCD consistently discovers communities that surpass the topology-semantic balance of the ground truth across multiple real-world networks. An empirical Pareto frontier analysis further validates that DBCD achieves a non-dominated solution, establishing it as a strong competitor among state-of-the-art methods. The source code of DBCD is available at https://github.com/wy980125/DBCD .}, archive = {J_ESWA}, author = {Yan Wang and Yupeng Liu and Xiaojie Sun and Jun Fu}, doi = {10.1016/j.eswa.2025.129487}, journal = {Expert Systems with Applications}, month = {3}, pages = {129487}, shortjournal = {Expert Syst. Appl.}, title = {DBCD: Deep balanced community detection via consensus-guided joint optimization in attributed networks}, volume = {298}, year = {2026}, }
• (2026). Network traffic forecasting with transfer learning-based algorithm for long continuous missing data. ESWA, 298, 129484. (www)

  @article{ , abstract = {Accurate network traffic forecasting is critical for power dispatch networks. Network traffic forecasting aims to use historical data to predict future network traffic trends. Different from other networks, the traffic data of the power dispatch network is mainly composed of the port traffic from routers and switches. However, network accidents in power enterprises can cause long periods of missing network traffic data, reducing the number of learning samples for network traffic prediction models and making the forecasting results unreliable. Due to the long periods of missing data, this paper uses transfer learning (TL) to impute missing data with the knowledge from a relevant task, which has ample samples. However, the imputation result contains complex source and target data characteristics. Therefore, this paper introduces the idea of frequency decomposition to decompose the imputation results into different sub-sequences through variational mode decomposition (VMD). Additionally, this paper uses long short-term memory (LSTM) networks to extract the potential features of decomposition results. Finally, this paper combines TL, VMD, and LSTM to design the TL-VMD-LSTM algorithm. The effectiveness of the proposed algorithm is validated using inflow and outflow traffic data from two State Grid Corp. of China networks. The results demonstrate that TL-VMD-LSTM has excellent generalization performance, with mean absolute percentage errors (MAPEs) of 0.380 % and 0.734 % for the Provincial access network and Information region network, respectively.}, archive = {J_ESWA}, author = {Yang Yang and Zhihao Chen and Yuchao Gao and Zijin Wang and Zhe Ding and Jinran Wu}, doi = {10.1016/j.eswa.2025.129484}, journal = {Expert Systems with Applications}, month = {3}, pages = {129484}, shortjournal = {Expert Syst. Appl.}, title = {Network traffic forecasting with transfer learning-based algorithm for long continuous missing data}, volume = {298}, year = {2026}, }
• (2026). High-dimensional multi-objective feature selection with niche-based binary differential evolution. ESWA, 298, 129478. (www)

  @article{ , abstract = {Feature selection is a critical step in machine learning and data mining, aiming to identify the most relevant features from a dataset to improve model performance while reducing computational costs. In high-dimensional data, as the dimensionality of data increases rapidly, feature selection faces an enormous search space, limiting the efficiency and effectiveness of traditional methods. To address these challenges, multi-objective optimization algorithms have emerged as a promising strategy for feature selection due to their ability to optimize multiple conflicting objectives simultaneously. We propose a niche-based binary differential evolution algorithm (MONBDE) for high-dimensional multi-objective feature selection. MONBDE enhances feature selection performance through several mechanisms: a niche-based binary differential evolution operator, redundant solution repair mechanism and an environmental selection strategy. In experiments, the proposed algorithm was compared with five advanced multi-objective optimization algorithms and tested on 15 benchmark datasets using three common metrics. Experimental results show that the MONBDE algorithm outperforms comparative algorithms in terms of classification accuracy and feature subset size across most datasets. The proposed strategy effectively eliminates redundant and irrelevant solutions in feature selection, leading to a significant improvement in model classification performance.}, archive = {J_ESWA}, author = {Xuezhi Yue and Xiang Zuo and Pengfei Ling and Chao Xiong and Hu Peng and Yuan Zeng}, doi = {10.1016/j.eswa.2025.129478}, journal = {Expert Systems with Applications}, month = {3}, pages = {129478}, shortjournal = {Expert Syst. Appl.}, title = {High-dimensional multi-objective feature selection with niche-based binary differential evolution}, volume = {298}, year = {2026}, }
• (2026). DFCNet: Dual-path fusion with intra-slice features and cross-slice constraints for cervical cancer CTV segmentation. ESWA, 298, 129477. (www)

  @article{ , abstract = {Accurate and rapid segmentation of the clinical target volume (CTV) is essential for cervical cancer radiotherapy. However, due to the soft boundaries of CTV, complex connections with surrounding tissues, and high interpatient variability, existing deep learning methods still face significant challenges, particularly for image slices that lack clear boundary information or where applicators are distant from CTV edges. Hence, we introduce DFCNet, a dual-path fusion network with cross-slice consistency constraints for CTV segmentation. The first path employs a dual-stream intra-slice feature encoding module to capture local and inter-regional details, thereby refining boundary delineation amidst the complex interplay with adjacent tissues. The second path integrates a cross-slice consistency constraint module to address soft boundaries and high interpatient variability, while ensuring the coherence and smoothness of the segmentation results. A feature fusion and decoding module combines semantic features from both paths, improving CTV region accuracy. Tests on 432 cervical cancer brachytherapy cases show DFCNet outperforms eighteen state-of-the-art segmentation methods, with Dice score improvements over two percentage points. The second path and feature fusion module can enhance other U-Net-based models, boosting their CTV segmentation performance. DFCNet excels in high-precision CTV segmentation, particularly for challenging slices, demonstrating its potential to improve cervical cancer radiotherapy accuracy, efficiency, and patient outcomes.}, archive = {J_ESWA}, author = {Mingxu Huang and Deyu Sun and Chaolu Feng and Ming Cui and Dazhe Zhao and Yuhua Gao}, doi = {10.1016/j.eswa.2025.129477}, journal = {Expert Systems with Applications}, month = {3}, pages = {129477}, shortjournal = {Expert Syst. Appl.}, title = {DFCNet: Dual-path fusion with intra-slice features and cross-slice constraints for cervical cancer CTV segmentation}, volume = {298}, year = {2026}, }
• (2026). Competitive e-commerce platforms’ data provision and pricing strategies with different attribution behaviors of users. ESWA, 298, 129466. (www)

  @article{ , abstract = {In the era of data, numerous e-commerce platforms are actively collecting consumer data to obtain business insights. However, consumers and merchants on platforms exhibit diverse attribution behaviors, including single-homing and multi-homing (access to only one platform/multiple platforms), not only affecting the platform’s market scale but also complicating their data provision strategies and data pricing strategies. Inspired by this practice, this paper considers varying attribution behaviors and studies the data operation strategies of competitive platforms. By constructing a two-period game model, we capture the entire process of platform’s data collection and provision, and solve the equilibrium decisions by reverse solution method. This research aims to identify the impact of attribution behaviors on platforms’ data strategies, thereby filling the gap in analyzing this issue from the perspective of two-sided platforms. Results show that when both groups of users (consumers and merchants) are multi-homing, platform facing higher operational costs may benefit more from implementing low data provision but high data pricing strategy, while more cost-efficient platform may choose the opposite strategy. This strategy is still applicable when only one group of users (consumers) becomes single-homing. However, once both groups of users are single-homing, both platforms’ strategies will change. Specifically, when merchant’s cross-side network effect (CNE) intensity is relatively low (high), compared with less cost-efficient platform, platform enjoying cost efficiencies should provide more (less) data at a relatively high (low) price. Moreover, platforms should cautiously provide data when both groups of users are single-homing, as it may hurt profits.}, archive = {J_ESWA}, author = {Wei Chen and Yijia Hu and Ronghua Sui and Zili Guan and Yi Liu}, doi = {10.1016/j.eswa.2025.129466}, journal = {Expert Systems with Applications}, month = {3}, pages = {129466}, shortjournal = {Expert Syst. Appl.}, title = {Competitive e-commerce platforms’ data provision and pricing strategies with different attribution behaviors of users}, volume = {298}, year = {2026}, }
• (2026). Emergency logistics planning through optimizing the multi-trip vehicle routing with time windows and limited trip duration. ESWA, 298, 129465. (www)

  @article{ , abstract = {Although the multi-trip vehicle routing problem with time windows (MTVRPTW) and its variants have been extensively studied, their application in natural disaster contexts remains underexplored. This study addresses this gap by developing a model and solution algorithm for the MTVRPTW with limited trip duration (MTVRPTW-LD), tailored to emergency supplies distribution in the early post-disaster phase. First, we replace the service-dependent loading time in traditional models with service-dependent unloading time and formulate an MTVRPTW-LD model to minimize total operational time, encompassing travel, service, and unloading times, based on the characteristics of emergency supplies distribution. Furthermore, a more practical method for calculating travel time is proposed to enhance the model’s applicability. Subsequently, a branch-and-price algorithm is designed to solve the MTVRPTW-LD model, in which the cumulative relative deprivation cost (CRDC) is introduced to improve equity in emergency supplies distribution. Finally, we conduct numerical experiments on Solomon instances and test instances generated based on emergency scenarios. The results show that, in the test instances, incorporating CRDC can improve the equity by up to 34.3 %.}, archive = {J_ESWA}, author = {Longfei Fan and Zhongming Wu and Zaiwu Gong and David Z.W. Wang}, doi = {10.1016/j.eswa.2025.129465}, journal = {Expert Systems with Applications}, month = {3}, pages = {129465}, shortjournal = {Expert Syst. Appl.}, title = {Emergency logistics planning through optimizing the multi-trip vehicle routing with time windows and limited trip duration}, volume = {298}, year = {2026}, }
• (2026). Regularity model-driven large-scale multi-objective evolutionary algorithm based on dual-information offspring reproduction strategy. ESWA, 298, 129460. (www)

  @article{ , abstract = {The Pareto set (PS) of a continuous multi-objective optimization problem exhibit a distribution along a low-dimensional manifold structure. This regularity property significantly contributes to generating high-quality offspring in large-scale multi-objective evolutionary algorithms (LSMOEAs). However, conventional regularity model-based algorithms face several challenges when dealing with large-scale multi-objective optimization problems (LSMOPs), including high computational costs for modeling, difficulty in capturing the true PS structure, and neglecting individual directional information. To address these challenges, we propose a dual-information offspring reproduction strategy that considers both the distribution information of the population and the directional information of the outstanding individuals. Specifically, this strategy comprises a sampling approach based on an augmented regularity model specifically designed for LSMOPs. Leveraging this model, we explore and exploit the decision space to sample a promising set of solutions. Additionally, the strategy also involves a search method based on competitive learning among individuals. By assigning a positive evolutionary direction to losing solutions, we update the losing solutions to generate high-quality offspring. We continuously refine the proposed regularity model to approximate the true PS more closely. In extensive experiments on large-scale multi-objective benchmark functions, we compare our algorithm with eight state-of-the-art algorithms. The results demonstrate that our approach excels in handling LSMOPs.}, archive = {J_ESWA}, author = {Ying Wu and Ziliang Du and Gonglin Yuan and Zhenzhou Tang and Ferrante Neri and Yaqing Hou}, doi = {10.1016/j.eswa.2025.129460}, journal = {Expert Systems with Applications}, month = {3}, pages = {129460}, shortjournal = {Expert Syst. Appl.}, title = {Regularity model-driven large-scale multi-objective evolutionary algorithm based on dual-information offspring reproduction strategy}, volume = {298}, year = {2026}, }
• (2026). GlobalCLIP: Zero-shot manufacturing anomaly detection with adaptive self-cyclic emsemble learning. ESWA, 298, 129448. (www)

  @article{ , abstract = {Accurately identifying product defects and process anomalies in manufacturing processes is a critical task for product quality and system stability. Although the existing unsupervised anomaly detection methods do not require the annotation of anomalies, they are difficult to deal with the zero-shot scenario faced by multi-variety and small-batch production where there is no available data. In addition, many zero-shot detection algorithms need to represent the image features in tensor form with multiple local feature vectors, and then measure each local feature to infer the overall anomaly of the object. In this paper, we propose GlobalCLIP, a novel approach for zero-shot anomaly detection using only global feature vectors to enhance performance. Specifically, we use CLIP model to aggregate the global features, and design two kinds of adaptive modules from the error level and uncertainty level to realize the series integration of different discriminant models. The adaptive modules encourage the model to learn both normal and abnormal patterns with different granularity, and the self-cyclic training progressively improves model performance. Experiments show that compared to many unsupervised/weakly supervised methods, the performance of GlobalCLIP maintains its advantage even without known samples, and achieves significant improvement over available zero-shot methods.}, archive = {J_ESWA}, author = {Haoyuan Shen and Enrico Zio and Jiawei Xiong and Yizhong Ma}, doi = {10.1016/j.eswa.2025.129448}, journal = {Expert Systems with Applications}, month = {3}, pages = {129448}, shortjournal = {Expert Syst. Appl.}, title = {GlobalCLIP: Zero-shot manufacturing anomaly detection with adaptive self-cyclic emsemble learning}, volume = {298}, year = {2026}, }
• (2026). Key performance indicator-related process monitoring for irregular scenarios with incomplete data. ESWA, 298, 129440. (www)

  @article{ , abstract = {Modern industries exhibit irregular characteristics due to factors such as mode transitions, incomplete data and outliers. Accurate monitoring of key performance indicators (KPIs) in irregular processes is essential for improving product quality and reducing scrap rates. This paper proposes a novel KPI-related process monitoring method that leverages the multiple kernel learning (MKL) technique, designed specifically for irregular scenarios with incomplete data. First, a novel MKL-based nonlinear matrix completion is proposed that utilizes a hierarchical strategy-based algorithm to estimate the missing values in incomplete data and the linear coefficients of multiple kernels. In addition, the corresponding convergence analysis is given. Based on the estimated completed data matrix, a novel MKL-based feature correlation analysis is proposed for indirect prediction of KPIs. Two statistics are established for detecting KPI-related and KPI-unrelated faults, respectively. A numerical case and an industrial example demonstrate that the proposed method not only accurately identifies the missing data, but also effectively detects the KPI-related faults.}, archive = {J_ESWA}, author = {Yanyu Chen and Hao Ma and Yan Wang and Xiang Liu}, doi = {10.1016/j.eswa.2025.129440}, journal = {Expert Systems with Applications}, month = {3}, pages = {129440}, shortjournal = {Expert Syst. Appl.}, title = {Key performance indicator-related process monitoring for irregular scenarios with incomplete data}, volume = {298}, year = {2026}, }
• (2026). A novel approach to anomaly detection in critical CPSs: The ERXAI-MS algorithm. ESWA, 298, 129437. (www)

  @article{ , abstract = {At present, practical machine learning methodologies are extensively employed for the automation of detecting the intrusion available in the network. In key infrastructure scenarios involving communication strategies, the interplay between different industrial control systems and the inherent connection to the Internet environment through the Internet of Things renders them vulnerable to cyber threats. Considering the substantial volume of network traffic within critical Cyber-Physical Systems, conventional machine-learning approaches utilized for detecting anomalies prove to be ineffective. Hence, newly designed machine learning methods, with a focus on deep learning, are demonstrating effective applications in identifying and categorizing anomalies on both network and individual device levels. This article introduces an innovative Ensemble Random Explainable Artificial Intelligence Meerkat Search algorithm designed for the identification of cyber threats. To augment the effectiveness of the suggested method, it employs a dual-step process for the detection of network irregularities. During the initial phase, the approach involves data pre-processing and dimensionality reduction through the application of Kernel Principal Component Analysis to select the most suitable features. In the subsequent stage, the novel Ensemble Random Explainable Artificial Intelligence Meerkat Search algorithm is employed for classification. The effectiveness of the approach presented in this study is evaluated on diverse datasets, encompassing information collected within the Internet of Things context, specifically IoT-23 and LITNET-2020 datasets. The findings of the assessment of the suggested method are deliberated upon, including the examination of statistical significance and a comparative analysis with contemporary approaches in the field of network anomaly detection. Evaluations confirmed this robust model attained 98.56% accuracy, 97.78% precision, 98.2% F1-score, and produced less FPR of 1.55%.}, archive = {J_ESWA}, author = {Prabakeran Saravanan and Annamalai Balaji and Hemalatha Murugan and Manickam Muruganantham and Indumathi Varadharajan}, doi = {10.1016/j.eswa.2025.129437}, journal = {Expert Systems with Applications}, month = {3}, pages = {129437}, shortjournal = {Expert Syst. Appl.}, title = {A novel approach to anomaly detection in critical CPSs: The ERXAI-MS algorithm}, volume = {298}, year = {2026}, }
• (2026). Improved DQN-based recommender system on three-way decision. ESWA, 298, 129431. (www)

  @article{ , abstract = {Recommendation systems, which utilize algorithms and data analysis to provide personalized suggestions to users, have become an indispensable part of modern life. However, traditional recommendation algorithms face challenges such as the cold start problem, lack of diversity, and limited scalability. Reinforcement learning (RL), particularly deep reinforcement learning (DRL), emerges as a promising solution to these problems by allowing agents to learn optimal strategies through interaction with their environment. Nevertheless, as the scale of data increases, RL-based recommendation systems often struggle to achieve a good balance between exploration and exploitation, impacting the overall performance of the algorithms. In this paper, we propose a reinforcement learning-based recommendation algorithm enhanced by a three-way decision (3WD) framework to address the exploration-exploitation balance challenge. The 3WD algorithm, rooted in rough set theory, categorizes decision outcomes into acceptance, rejection, and uncertainty regions. By applying 3WD in the action selection process of RL, we optimize the trade-off between exploration and exploitation, thereby improving the quality and computational efficiency of recommendations. Additionally, we introduce a dynamic threshold adjustment mechanism to adaptively refine the decision boundary during the action selection process in reinforcement learning, further enhancing the algorithm’s performance. Using the MovieLens dataset as a foundation, we conduct extensive experiments with several randomly generated data sets to evaluate the proposed method. Our results demonstrate that the 3WD-based RL algorithm outperforms traditional methods, such as epsilon-greedy and Softmax, in terms of runtime, recommendation accuracy, and error rate. Notably, the dynamic threshold adjustment model exhibits greater stability and surpasses static methods in recommendation success rates. These findings highlight the effectiveness of combining 3WD with RL in recommendation systems, providing a powerful and efficient solution to the challenges faced by traditional methods. Finally, we analyze the limitations of the model based on the experimental results and propose avenues for future research.}, archive = {J_ESWA}, author = {Zian Chen and Bao Qing Hu}, doi = {10.1016/j.eswa.2025.129431}, journal = {Expert Systems with Applications}, month = {3}, pages = {129431}, shortjournal = {Expert Syst. Appl.}, title = {Improved DQN-based recommender system on three-way decision}, volume = {298}, year = {2026}, }
• (2026). IntegralGP: Volumetric estimation of subterranean geochemical properties in mineral deposits by fusing assay data with different spatial supports. ESWA, 298, 129429. (www)

  @article{ , abstract = {This article presents an Integral Gaussian Process (IntegralGP) framework for volumetric estimation of subterranean properties in mineral deposits. It provides a unified representation for data with different spatial supports, which enables blasthole geochemical assays to be properly modelled as interval observations rather than points. This approach is shown to improve regression performance and boundary delineation. A core contribution is a description of the mathematical changes to the covariance expressions which allow these benefits to be realised. The gradient and anti-derivatives are obtained to facilitate learning of the kernel hyperparameters. Numerical stability issues are also discussed. To illustrate its application, an IntegralGP data fusion algorithm is described. The objective is to assimilate line-based blasthole assays and update a block model that provides long-range prediction of Fe concentration beneath the drilled bench. Heteroscedastic GP is used to fuse chemically compatible but spatially incongruous data with different resolutions and sample spacings. Domain knowledge embodied in the structure and empirical distribution of the block model must be generally preserved while local inaccuracies are corrected. Using validation measurements within the predicted bench, our experiments demonstrate an improvement in bench-below grade prediction performance. For material classification, IntegralGP fusion reduces the absolute error and model bias in categorical prediction, especially instances where waste blocks are mistakenly classified as high-grade.}, archive = {J_ESWA}, author = {Anna Chlingaryan and Arman Melkumyan and Raymond Leung}, doi = {10.1016/j.eswa.2025.129429}, journal = {Expert Systems with Applications}, month = {3}, pages = {129429}, shortjournal = {Expert Syst. Appl.}, title = {IntegralGP: Volumetric estimation of subterranean geochemical properties in mineral deposits by fusing assay data with different spatial supports}, volume = {298}, year = {2026}, }
• (2026). Deep multi-view least squares support vector machine with consistency and complementarity principle based on cross-output knowledge transfer. ESWA, 298, 129406. (www)

  @article{ , abstract = {This article presents a Deep Multi-view Least Squares Support Vector Machine with Consistency and Complementarity Principles based on Cross-Output Knowledge Transfer (MDCTM), which has four distinctive features: 1) It integrates the idea of deep stacking architecture, which is the first attempt to use transfer learning to form deep architectures in multi-view learning. It can enhance the ability to handle complex problems. Starting from the second layer, it incorporates extra input attributes that consider the predictions made by all preceding layers, effectively revealing the manifold structure of the original data. 2) Each layer follows the consistency and complementarity principles, which can fully excavate the information in multi-view data. In each layer, the model is solved by an alternating optimization strategy. 3) Cross-output knowledge transfer leverages predictions from earlier layers to improve the learning of subsequent ones, which can improve the classification performance of the model. Additionally, the extent of cross-output knowledge transfer between sequential layers can be assessed autonomously and effectively by utilizing a fast leave-one-out cross-validation method. 4) The model allows random assignment of model parameters in each layer, such as weights and kernel widths, boosting learning speed. Numerical experiments demonstrate the model’s effectiveness and efficiency.}, archive = {J_ESWA}, author = {Shuangrui Jia and Sijie Liang and Ziyi Mo and Chunxiao Liu and Huiru Wang and Chen Chen}, doi = {10.1016/j.eswa.2025.129406}, journal = {Expert Systems with Applications}, month = {3}, pages = {129406}, shortjournal = {Expert Syst. Appl.}, title = {Deep multi-view least squares support vector machine with consistency and complementarity principle based on cross-output knowledge transfer}, volume = {298}, year = {2026}, }
• (2026). A study on the modeling of long-term trend and short-term fluctuation: Fourier-enhanced adaptive koopman operator for carbon emission forecasting. ESWA, 298, 129388. (www)

  @article{ , abstract = {The Koopman operator provides a new way to model complex data patterns, revealing the intrinsic dynamics of time series from a dynamical system perspective. Despite its potential, the Koopman operator has received limited attention in time series forecasting, particularly in addressing these complex, real-world challenges such as carbon emission dynamics characterized by nonlinearity, non-stationarity, and multi-scale coupling effects. To this end, this study proposes a novel forecasting paradigm, Fourier-Enhanced adaptive Koopman operator for carbon emission forecasting (F-KOCE). This approach conceptually extends traditional Koopman frameworks by embedding a spectral-decoupled time series representation into a dual Koopman learning structure, which enables the model to linearize nonlinear dynamics across multiple time scales in a theoretically grounded and practically adaptive manner. By integrating Fourier filter decomposition into Koopman operator theory, F-KOCE separates raw emissions into long-term trends and short-term fluctuations while achieving global linearization of system dynamics. A learnable Koopman operator captures intrinsic temporal structures, while a multi-granularity adaptive weight learning strategy enhances resilience against data variability. To further improve robustness, we introduce an adaptive residual fusion structure for block-level feature compression, noise suppression, and cross-scale information fusion. Additionally, the effective Trend Corrector mechanism dynamically modulates the influence of trend and fluctuation components, refining predictive accuracy. Beyond point forecasting, the framework is also extended to interval forecasting, providing uncertainty-aware predictions. Extensive experiments conducted on 36 Carbon Monitor datasets across six regions and six sectors demonstrate the superiority of F-KOCE over advanced existing models across multiple evaluation metrics. These results confirm the framework’s efficacy in capturing high-dimensional emission dynamics and underscore the potential of Koopman operator theory in carbon forecasting. By offering a robust, interpretable, and data-driven approach, F-KOCE provides valuable insights for climate policy formulation.}, archive = {J_ESWA}, author = {Jinxing Che and Wei Dong and Qian Sun and Yuhua Zhang}, doi = {10.1016/j.eswa.2025.129388}, journal = {Expert Systems with Applications}, month = {3}, pages = {129388}, shortjournal = {Expert Syst. Appl.}, title = {A study on the modeling of long-term trend and short-term fluctuation: Fourier-enhanced adaptive koopman operator for carbon emission forecasting}, volume = {298}, year = {2026}, }
• (2026). A cooperative co-evolutionary algorithm with core-based grouping strategy for large-scale 0–1 knapsack problems. ESWA, 298, 129364. (www)

  @article{ , abstract = {The 0–1 knapsack problem (KP) is a well-known combinatorial optimization problem with wide real-world applications. While evolutionary algorithms have demonstrated promise in solving 0–1 KPs, their performance deteriorates as the problem dimension increases. Cooperative co-evolution (CC) is an algorithmic framework based on a divide-and-conquer strategy, which has been used in solving large-scale optimization problems. Inspired by the similarity between item grouping in the 0–1 KP and decomposition strategies in CC, this paper proposes a novel grouping strategy that uses the position information of break items and profit-to-weight ratio to solve large-scale 0–1 KP. The strategy aims to divide the large-scale 0–1 KP into multiple subproblems, thus having a reduced search space for each subproblem. To enhance population diversity and search efficiency, the profit-to-weight ratio is used to generate an initial elite population. Additionally, to obtain the complete solution for the original large-scale KP, a subgroup merging method is designed to accelerate convergence and further improve population diversity. A three-phase repair operator is developed to fix infeasible solutions directly to create more feasible solutions. The resulting cooperative co-evolutionary algorithm is compared with ten state-of-the-art algorithms for solving 0–1 KPs with variables ranging from 100 to 5,000, including EAs, CC-based approaches, and a deep reinforcement learning method. Experimental results show that the proposed algorithm exhibits higher solution accuracy and faster convergence than other competing algorithms. The CC framework takes considerably less running time than high-performing algorithms, providing an overall novel approach for solving large-scale 0–1 KPs.}, archive = {J_ESWA}, author = {Xiaotong Li and Shuwei Zhu and Wei Fang and Kalyanmoy Deb}, doi = {10.1016/j.eswa.2025.129364}, journal = {Expert Systems with Applications}, month = {3}, pages = {129364}, shortjournal = {Expert Syst. Appl.}, title = {A cooperative co-evolutionary algorithm with core-based grouping strategy for large-scale 0–1 knapsack problems}, volume = {298}, year = {2026}, }
• (2026). Life cycle cost reliability assessment for strategic real estate decision-making. ESWA, 298, 129329. (www)

  @article{ , abstract = {Many real estate projects prioritize minimizing initial development costs while often overlooking the long-term financial implications of their decisions. This short-term focus frequently leads to increased operational, maintenance, and renewal expenses, ultimately reducing overall profitability. Life Cycle Costing (LCC) provides a comprehensive approach to evaluating total project costs over time; however, its adoption remains limited due to challenges such as data constraints, uncertainty about future cost savings, and the lack of standardized performance measurement tools. To tackle these issues, this paper proposes a structured LCC reliability assessment model designed for real estate decision-makers. The model systematically identifies and analyzes key cost factors across all project phases, including construction, operation, renewal, maintenance, and end-of-life, while integrating technical, economic, environmental, and social dimensions. A structured survey was employed to quantify and prioritize these cost factors, facilitating the development of category-specific LCC models and a standardized evaluation framework. Additionally, a benchmarking scale was created to measure the reliability of input factors. Although this study emphasizes the reliability dimension, it establishes a foundation for the future integration of an optimization module to enhance decision-making and maximize life cycle cost efficiency. The proposed model has been automated to improve usability and accessibility, allowing stakeholders to make informed investment decisions that promote long-term financial sustainability in real estate development.}, archive = {J_ESWA}, author = {Elin A. Eldars and Amin A. Sorour}, doi = {10.1016/j.eswa.2025.129329}, journal = {Expert Systems with Applications}, month = {3}, pages = {129329}, shortjournal = {Expert Syst. Appl.}, title = {Life cycle cost reliability assessment for strategic real estate decision-making}, volume = {298}, year = {2026}, }
• (2026). Balancing forecast accuracy and switching costs in online optimization of energy management systems. ESWA, 298, 129305. (www)

  @article{ , abstract = {This study investigates the integration of forecasting and optimization in energy management systems, focusing on how switching costs, penalties incurred from frequent operational adjustments, affect the balance between forecast accuracy and stability in online decision-making. We develop a theoretical framework analyzing Fixed Horizon Control (FHC) algorithms under switching costs, deriving performance bounds that reveal trade-offs between commitment periods and forecast properties. We introduce a novel Scenario Distribution Change (SDC) metric for measuring temporal consistency in probabilistic forecasts. The framework is validated through empirical evaluation using a real-world battery scheduling case study based on the CityLearn 2022 challenge, comparing deterministic and stochastic optimization approaches across different commitment periods. Theoretical analysis reveals that switching costs create a U-shaped relationship between commitment period and performance, with optimal commitment depending on forecast stability. Empirical results demonstrate that switching costs significantly alter the accuracy-stability trade-off: while traditional approaches favor frequent updates (1-hour commitment), incorporating switching costs makes longer commitment periods (3+ hours) optimal when combined with stable forecasts. Stochastic optimization with scenario averaging reduces forecast error sensitivity by up to 2.9 % in grid costs compared to deterministic approaches. This work contributes the first theoretical bounds linking forecast stability to switching costs in energy systems, the SDC metric for evaluating probabilistic forecast stability, empirical evidence that longer commitment periods can outperform frequent updates under switching costs, and practical guidelines showing that forecast stability should be factored into decision-making frameworks for energy management systems in the presence of switching costs.}, archive = {J_ESWA}, author = {Evgenii Genov and Julian Ruddick and Christoph Bergmeir and Majid Vafaeipour and Thierry Coosemans and Salvador García and Maarten Messagie}, doi = {10.1016/j.eswa.2025.129305}, journal = {Expert Systems with Applications}, month = {3}, pages = {129305}, shortjournal = {Expert Syst. Appl.}, title = {Balancing forecast accuracy and switching costs in online optimization of energy management systems}, volume = {298}, year = {2026}, }
• (2026). A hybrid bee colony algorithm for crack repair trajectory planning based on focal attention guided lightweight segmentation. ESWA, 298, 129267. (www)

  @article{ , abstract = {Robotic arm trajectory planning of crack repair is critical for automated road maintenance. However, existing crack repair face two main challenges: loss of trajectory edge information and redundant planned distances. This study introduces an automated pavement crack repair system that integrates a lightweight crack segmentation model (Lightweight Focal Modulation, LFM-Net) and a repair trajectory planning algorithm (Fixed Neighborhood Search-Artificial Bee Colony, FNS-ABC). Specifically, LFM-Net incorporates conformer-based focal modulation attention (CFMA), enhancing the detailed information during the decoding phase. Additionally, the FNS-ABC enhances the ABC algorithm by incorporating a fixed neighborhood search strategy, effectively reducing redundant planning paths. The system is executed using a self-developed robotic arm with an edge computing unit. Extensive testing in three typical road scenarios-independent cracks, intersection cracks, and complex cracks-demonstrated that the system achieved a mean Intersection over Union (mIoU) of 83.93 %. Finally, the system exhibited an idle trajectory of 79.51 mm when addressing complex cracks, highlighting its superior performance in repair trajectory planning.}, archive = {J_ESWA}, author = {Jianqi Zhang and Xu Yang and Wei Wang and Yuhang Zhao and Hainian Wang and Yixue Chen}, doi = {10.1016/j.eswa.2025.129267}, journal = {Expert Systems with Applications}, month = {3}, pages = {129267}, shortjournal = {Expert Syst. Appl.}, title = {A hybrid bee colony algorithm for crack repair trajectory planning based on focal attention guided lightweight segmentation}, volume = {298}, year = {2026}, }
• (2026). Deep-learning based big data analysis for developing a smart supply chain for increased efficiency. ESWA, 298, 129246. (www)

  @article{ , abstract = {Big data analysis (BDA) in supply chain management (SCM) is receiving growing attention in the present business environment. This is due to the fact that BDA has a wide range of applications in SCM, including customer behaviour analysis, trend analysis, and demand prediction. The increase in information volume has caused the efficiency and effectiveness of traditional procedures to decline, considering this, researchers have developed techniques that have a high capacity to investigate and comprehend vast amounts of data due to the limitations of these tactics in dissecting and interpreting a lot of information. This study represents a hybrid paper that combines a systematic literature review, a methodological proposal using BP neural networks. The main objective of this paper is to recognize the uses of deep learning in SCM. By fostering a calculated system, this paper recognizes the commitments of deep learning strategies in choosing and sectioning providers, foreseeing store network gambles, and assessing requests and deals, creation, stock administration, transportation and circulation, manageable turn of events, and roundabout economy. The novelty in this paper is the Backpropagation (BP) neural networks with big data-driven demand forecasting in supply chains. This method can improve the accuracy of demand forecasting in supply chain management. The study includes a thorough survey of the applications of predictive BDA in SC request gauging. The review highlighted the BDA methodologies used for production network request estimation and comparatively classified them. We collected and analysed these studies as tactics and methodologies for the popular forecast. Seven standard tactics were selected and studied, along with their benefits and drawbacks. Finally, the box-cox transformation representation over years in which for the year 2011–01, it starts with a box-cox value of 9.7 and it inclined till 2011–06 and then declined very exponentially in 2012–07 at 9 and then it keeps on incrementing and reached at 11 at the year 2013–07. Then from 2014 to 2015, the pattern didn’t lower below the box-cox value 10.}, archive = {J_ESWA}, author = {Sreekumar Narayanan and Sudhir Ramadass and K. Thilagavathi and Rajiv Kumar}, doi = {10.1016/j.eswa.2025.129246}, journal = {Expert Systems with Applications}, month = {3}, pages = {129246}, shortjournal = {Expert Syst. Appl.}, title = {Deep-learning based big data analysis for developing a smart supply chain for increased efficiency}, volume = {298}, year = {2026}, }